APPROXIMATE REASONING ABOUT TEMPORAL CONSTRAINTS … · Shashi SHEKHAR N 90/13/TM Assistant...
Transcript of APPROXIMATE REASONING ABOUT TEMPORAL CONSTRAINTS … · Shashi SHEKHAR N 90/13/TM Assistant...
"APPROXIMATE REASONING ABOUT TEMPORALCONSTRAINTS IN REAL TIME PLANNING AND
SEARCH"
bySoumitra DUTTA*
andShashi SHEKHAR**
N° 90/13/TM
Assistant Professor of Information Systems, INSEAD, Boulevardde Constance, 77305 Fontainebleau, France
* * Department of Computer Science, University of Minnesota,Minneapolis, U.S.A.
Printed at INSEAD,Fontainebleau, France
APPROXIMATE REASONING ABOUT TEMPORAL CONSTRAINTS
IN
REAL TIME PLANNING AND SEARCH
Soumitra Dutta Shashi ShekharTechnology Management Area Dept. of Computer Science
INSEAD
University of MinnesotaFontainebleau, France 77305
Minneapolis, USA
Abstract
Artificial Intelligence (AI) systems are being increasingly applied to
challenging real time problems. Real time AI systems typically have to
operate under stringent temporal constraints. These temporal constraints
can be either explicit (e.g., a system may have to respond within a
given deadline) or implicit (e.g., there may be no explicit deadlines but
optimal response times may be desired). This paper is concerned with
approximate reasoning about temporal constraints present either
explicitly or implicitly in real time systems. A graph search procedure is
used as the conceptual representation for real time reasoning and
planning. The cost of a response consists of the cost to plan a solution
and the cost to execute the chosen solution. There is an intimate
tradeoff between these two costs. Algorithms for approximate reasoning
about such tradeoffs are discussed in this paper. In particluar, a
heuristic algorithm called NORA is proposed for determining the tradeoff
between planning and execution costs and a formal proof is given for.
the performance bound obtained. The application of NORA to the domains
of semantic query optimization and the A* heuristic search algorithm is
described.
1
2
1 Introduction
As the application of AI systems evolves from an art to an engineering
science, more challenging applications shall be addressed. Some of the most
challenging and interesting applications can be found in real-time domains. In
real-time domains, the external world is dynamic and potentially ever-
changing. Three major phases can be identified in the solution structure:
[1] Data Collection : This phase refers to the collection of data about the
external world either through individual observations or from the
outputs of various sensors.
[2] Planning & Reasoning : This phase includes all planning and reasoning
that is done on the data (collected during the data collection phase)to obtain a solution. The terms planning and reasoning are used here
in a very general sense to include all different manipulations of data
used to obtain a solution and may include one or more of a diverse
range of techniques such as table lookups and heuristic rule-based
methods.
[3] Execution : This is the final phase during which the solution obtained in
the earlier phase (of planning ) is executed In the real world.
Figure 1 about here
Typically, these three phases occur sequentially (as shown in figure 1), i.e.,
first some data is collected, then a solution is found and finally the solution is
executed. In general however, the divisions between these three different
phases are neither crisp nor universal. Thus It Is quite possible for these
various phases to overlap, e.g., some additional data may have to be collected
during the planning phase or some replanning may be necessary during the
execution phase. For simplicity, it may be assumed that the three phases occur
sequentially and do not overlap (as shown in figure 1).
3
Each of the three phases shown in figure 1 typically consume some time as
shown in figure 2. In figure 2, the time axis begins at some arbitrary point of
time, to, and the time periods consumed by each of the three phases are:
Data Collection: [ to, t1 3
Planning : [ t 1 , t2 3
Execution : [ t2, t3 3
Figure 2 about here
Let W(t) represent the state of the external dynamic world at time t. It can
be seen from figure 2, that the state of the world at the beginning of the
execution phase, W(t2 ) is potentially different from that at the beginning of
the data collection phase, W(to). Even if it is assumed (reasonably) that the
data used for planning correctly represents the state of the world at time t1,
the end of the data collection phase, it is useful to minimize the sum of the
planning and execution times in order to minimize the difference between the
state of the world, W(t 1 ), on which the planning is based and the states of
the world, W(t2 ) through W(t3 ), during which the solution is executed. This
would maximize the chances that the selected solution can actually be executed
(i.e., the world has not changed so as to prevent the execution of the chosen
solution) and would minimize the time necessary to observe the effects of the
executed solution. It is necessary to provide some capability for meta-level
reasoning about the relative durations of these three different phases. For
example, any excessive expenditure of time in the planning phase may offset
the gains (possibly obtained by the extra planning) in the execution time and
may also render the solution determined on the basis of data collected from
W(t 1 ) useless for execution in the world at time t 2, as W(t2 ) through W(t3 ) may
be drastically different from W(t 1 ). The situation is more complicated in the
case of deadlines.
Such kind of meta-level temporal constraints on planning and execution are
common in artificial intelligence (AI) systems operating in real-time situations
which typically have to respond within a certain deadline, or have optimal
response times (for planning and execution). For example in managing
4
defensive weapons against nuclear missiles, the system has to respond within
a few seconds or minutes [1]. On the other hand, in credit approval systems
[2], optimal response times are desired, but there are no hard deadlines.
Previous research in artificial intelligence has typically focussed solely on
solution techniques for use during the planning and reasoning phase and
there is considerable literature on a variety of related topics, such as
knowledge representation, rule-based systems, etc. Real time systems and its
related issues of meta-level temporal constraints of deadlines and optima
response times have been largely ignored by the artificial intelligence
community. Current AI systems are usually designed to demonstrate technology
applications, and most planning and problem-solving algorithms are not
structured to meet real time (temporal) constraints. For example, planning
methods based on A* [3] and IDA* [4] may potentially take exponential time for
problem-solving. Within a given deadline to solve a certain problem, the
system may not produce any solution (complete or partial). Scaling them up-
with real world data and adequate knowledge bases, would amplify their
performance problems, including their inability to meet real time constraints
[5]. Korf [6, 7] has worked on planning and search algorithms to meet strict
deadlines which guarantee a partial solution within a fixed deadline. However,
his work has ignored execution time and essentially spent all available time for
planning. There has been no previous work on planning algorithms which
take into account the times spent on the data collection, planning and
execution phases of the total solution structure (figure 2). This paper
addresses this important aspect of real time problem solving and develops
algorithms that perform meta-level reasoning about temporal constraints on the
relative durations of the data collection, planning and execution times in real
time planning.
2 Motivation
Figure 3 about here
There are two important classes of real time constraints (see figure 3):
5
[1] Deadlines : This is the case wherein a real time system has to produce a
response within a fixed deadline. For example, in some chess
tournaments, moves have to made within a fixed time limit. The deadline
problem is hard and many such problems (e.g., scheduling with
deadlines) are NP-complete [8]. Providing a guaranteed optimal response
within a fixed deadline is not possible always and usually at most a
partial solution or a coarse solution within the given time constraint can
be guaranteed. For many problems, partial solutions are acceptable. An
example of such a problem is robot path planning, where a robot is
required to move as far as possible within the deadline when asked to
move from point A towards point B within a fixed time limit.
[2] Optimal response times : This is the case where the desire is to produce
and execute a solution within the least amount of time. Ignoring data
collection time, the response time is the sum of planning time amp
execution time, and the application requires an optimal response time. As
shown in figure 4, in general, the optimal response time is not achieved
by solely minimizing execution time, as then planning time increases to
offset the gains. The American Express credit card transaction approva!
system [2] is an example of such a system. Here the system response
time should be as small as possible. Many on-line information retrieval
systems also fall under this class of real time systems.
Figure 4 about here
The above classification of real time systems has been mentioned in the
literature [9], which provides the following two common definitions of real
time:
[1] There is a strict time limit by which the system must have produced a
response, regardless of the algorithm employed [93.
[2] The system is predictably fast enough for use by the process being
serviced [10].
These definitions each emphasize different aspects of real time performance.
The first definition is representative of the deadline problem faced by many
6
real time systems. Currently, ad hoc techniques are used for making a system
produce a response within a specified time interval, and these methods suffer
from poor extensibility, brittleness and a lack of a formal proof of "reliable
real time performance" [9].
The second definition is less stringent on the hardness of the deadline it
places on the performance of the real time system and is representative of the
optimal response time problem described earlier. Consider the case of Ft user
querying a database from a terminal. The user types in a query ant; waits fc-
an answer. He imposes no strict deadline, but also does not want to be kept
waiting for an inordinate period of time. Thus the system should minimize its
response time so as to provia the best possible real time service. There are
several other applications where providing the optimal response time is of
overridin2 importance.
Ur1der temporal constraints of either deadlines or optimal response times, a
real time system must perform some meta level reasoning to determine the
relative amounts of times to be spent on each of the three solution phases
shown in figures 1 and 2. When faced with a deadline to plan and execute a
solution, a real time system has to stop planning at some time instant ahead of
the deadline such that it has sufficient time left to execute the chosen
solution. The tradeoffs between the relative durations of the various phases
becomes more complex under the constraint of providing optimal response
times. For example, a greater planning cost may possibly lead to a lower
execution cost, but the extra time spent planning may also lead to a potential
Increase in the response time as shown In figure 4. Thus it is crucial that a
real time system be able to decide upon the appropriate moment to stop
planning and start executing the best solution plan obtained till then. Korf [6]
has mentioned the existence of this tradeoff and agreed that
"in principle we could find algorithms that minimized total solution time
by balancing thinking time and action time".
This paper determines algorithms which consider this trade-off between
thinking time and action time and proves bounds on the optimality of the
results obtained. A general conceptual framework of planning (described in the
7
next section) is chosen for the exposition of the results.
3 Problem Definition
Though the data collection phase extends from to to t 1 in figure 2, it is
reasonable to assume that the data collected accurately reflects the state of
the world at time t 1 . This data about W(t1 ) is then used for the planning
phase. It can be assumed that the sensors used to collect data about the
external world provide all the data at time t 1 . The data collection phase can
be considered to occur around the instant, t 0 = t 1 , which simplifies the
problem of meta-level temporal reasoning down to the study of the tradeoff
between the planning and execution times. This simplification is quite
reasonable for many applications. For example, in some engineering situations,
there are sensors that continuously monitor some aspect of the external woric
(such as a voltmeter monitoring the voltage) and the processing unit may at
some instant sample the values of these sensors for use in further planning
and execution.
The tradeoff between planning and execution times was graphically
represented in figure 4. It is desirable that the meta-level reasoning be able
to stop planning and start executing as close to the optimal point (of minimum
response time) as possible. Before attempting to provide any principles to
guide such action, it is necessary to state clearly the conceptual model of
planning in the context of which such principles can be stated.
Planning is a very general term, with different connotations to different
research communities. Within AI, it generally refers to reasoning about actions.
A recent book by Wilkins [11] provides a good survey and introduction of
relevant issues in the area. Search is an important part of any planning
system and there are several good reviews of search strategies in the context
of planning [12, 13, 14]. Researchers such as Korf [15] have supported the
thesis that planning can be viewed as problem solving search. Wilkins [16] has
drawn the distinctions between searching a space of partial plans as opposed
to searching a space of world states, and searching through plan
specializations rather than through plan modifications. This paper considers a
very general description of the search space. Conceptually, the search space
8
can be modeled as a set of nodes with inter-connecting arcs. Each node is a
possible plan and has an associated execution time. Each arc represents either
plan specialization or plan modification. The planning time associated with each
node consists of the sum of the cost of planning at the node and the cost ot
reaching that node from the previous node. These costs are measured in some
common units (time). The cost of planning at a node includes all costs
associated with planning at that node, such as plan generation and pre-
condition checking. The cost of arc traversal includes the cost of generating
suitable plan refinements or plan modifications. The exact nature of the
division of costs between nodes and arcs is not crucial to the results
presented below. The conceptual idea important to the results is that planning
consists of a search in a graph of possible plans for the best possible plan
(i.e., the least cost plan) and that it is possible to keep track of total
planning cost (time) incurred at any stage. This view is very general and can
be applied to most current planning systems.
In this context, the problem addressed in this paper can be simply stated as:
To propose and study principles guiding the trade-offs between
planning and execution costs in real time systems
The aim is the determination of principles which help a real-time system to
perform meta-level reasoning to decide when to stop planning and when to
start executing the best solution found so far. The term best solution is
used in this paper to refer to the solution with the least execution cost. For
clarity and simplicity, this cost is equated to the execution time. Implicit in
this usage Is the simplification that the quality of the solution is being
measured primarily by the time taken for execution. This assumption is valid
only for certain domains, such as query optimization. In many other domains,
the quality of a solution is a multi-attribute vector, one component of which
is the execution time. However, the theoretical developments and analyses
developed here shall be equally valid for such situations as long as planning
and execution costs are expressed in some common units (some virtual units,
such as utils to express utility can be used).
The emphasis is also on cases where the planning and execution costs are
9
comparable, as only then is it important to consider their relative trade-offs.
If either the planning or execution cost is small, some simplifications are
possible. For example, if planning cost is small, it may be possible to search
the entire search space quickly and find the optimal solution to execute. On
the other hand, if execution cost is small, it is again possible to spend most of
the time on planning. This latter case is the usual case considered by
algorithms like A*, which typically ignore exeuction costs.
In general, several factors can complicate the determination of optimal policies
deciding the trade-off between planning and execution costs. Most real life
situations are extremely complex and the associated search spaces are very
large. Searching the entire search space to determine the best solution may
require prohibitively large amounts of time and thus may not be possible in
real time situations. Also, it is often not possible to apriori specify the goal
node, i.e., the solution with the least execution cost. This is in contrast to the
requirements of popular search algorithms such as A* (and even its real
time equivalent RTA* proposed by Korf [6, 7]), which require that the goal
node be specified apriori .
4 Review of Previous Research
Surprisingly, there has been little research in artificial intelligence on real
time search and planning. Simple blind search algorithms like depth first
search, breadth first search and depth first iterative deepening [4] are useful
with small search spaces and large deadlines. However, the size and complexity
of most search spaces faced by real time AI systems preclude the use of these
algorithms. Heuristic search algorithms like A* and IDA* are not very useful
for real time problem domains as they may take exponential time in producing
a solution. These search strategies may be able to guarantee an optimal
solution in the absence of limits on search time, but cannot function with the
constraint of deadlines and response times. The notion of a deadline limits the
search horizon and often necessitates the taking of an action before the
optimal solution is found.
Korf [6] has formulated variations on minmax search and A* to adapt to the
real time scenario. His algorithm for bounded look-ahead search in single
10
agent problems is called minmin search and it involves searching forward
from the current state to a fixed depth horizon determined by the deadline
and then applying the A* cost function of f(n) = g(n) + h(n), to the frontier
nodes. The minimum value is then backed up and a single move is made in the
direction of the minimum value. Making a single move at a time follows a
strategy of least commitment and allows a different second move to be
recommended by further search (caused by the dynamically changed
environment). He also proposes an analog to alpha-beta pruning that makes
the same decision as minmin search but by exploring fewer nodes.
Korf has further proposed a real time modification of A* called RTA* for
controlling the sequence of moves actually executed. The neighbors of the
current state are generated and a heuristic function, including look-aheac
search with alpha pruning is applied to each new state. The neighbor with the
minimum g+h value is chosen as the new current state, and the old current
state is stored in a table along with the second best g+h value, which is the
best value among the remaining children. This represents the best estimate of
the cost of finding the solution via the old current state from the perspective
of the new current state. Russell [17] has extended RTA* by adding meta-
greedy decision-theoretic search control to RTA*.
The real time algorithms proposed by Korf essentially deal with the problem of
deadlines in real time systems (see classification in figure 3). They ignore the
execution cost of any solution and essentially spend all available time (the
deadline) searching. This is not realistic in many situations where execution
time is appreciable.
The above-mentioned concerns about "thinking time and "action time in real
time situations is a special case of the more general issue of bounded
rationality [18] and computation under resource constraints. There has been
considerable research both within decision theory and conventional AI, on
normative bases for reasoning. Several researchers [18, 19, 20] have pointed
to deficiencies in using normative theories (i.e., probability theory and utility
theory) as a consistent axiomatic basis for complex, real world inferences. The
earliest discussion on the explicit integration of the costs of inference within
the framework of normative rationality was introduced by Good [21]. Good
11
defined type I rationality as inference that is consistent with the axioms of
decision theory without regard to the cost of inference. Type II rationality
was defined as behavior that takes into consideration the costs of reasonina.
Several researchers such as Horvitz [18, 20], Matheson [22], Watson and Brown
[23], and Lindley [24] have explored the integration of computation costs into
inference strategies in the spirit of the work by Good. However, their primary
emphasis has been on techniques for reformulating a base problem into one
that will be of greater value than a complete analysis would be, given
computational resource constraints. The trade-offs considered by them have
been different from that considered in this paper, e.g., Horvitz [19] has
studied the trade-off between the immediacy and precision of the results
and how it affects the selection of alternative computation strategies. A
criticism of these approaches is that they often require detailed apriori
knowledge of relative utilities and precisions. An expert may have difficulty in
articulating such numbers and they may not be available always.
An alternative approach has been taken by researchers such as Bonissone and
Halverson [25] who have focussed on pre-compiled knowledge about the
problem solving domain in order to control the search strategy and to
select/construct the most valuable inference strategy. Such an approach is
applicable only for well structured situations in which such detailed knowledge
is available apriori for compilation. The approach taken in this paper is
different in that it does not assume the availability of any such knowledge for
compilation.
5 Planning & Search With Qptimal Response Times
The trade-offs between planning and execution costs become significant while
trying to provide optimal response times in real time systems where the
planning and execution costs are comparable (figure 4). Without explicit apriori
knowledge about the domain characteristics (both static and dynamic), it is not
easy to devise algorithms that can always provide the optimal response time.
In general, such explicit knowledge is very difficult to obtain and probably
very unreliable, as the external world can change in a myriad number of
ways. Also, as the goal is often not known apriori , it is very difficult (in
general) to come up with a precise stopping rule. Under such circumstances,
12
it is best to try to devise algorithms that provide some heuristic principles to
guide the tradeoff between planning and execution costs. While good heuristics
are often useful, it is even better If bounds can be proved on the obtained
response. This section describes NORA, a simple and intuitively appealing
planning and search algorithm for providing near optimal response times in
real time systems.
5.1 NORA: A Near Optimal Resoonse-time Algorithm
Blind exhaustive searches like BFS or DFS are generally applicable to limited
domains. Most heuristic search algorithms, e.g., A*, require the ability to
recognize the goal node (i.e., the best plan). This is not always possible, (e.g.,
recognizing the cheapest query execution plan in semantic query optimizatior
[26]) and one needs other heuristics to control the search.
Given a search algorithm, a real time system has to make a crucial decision of
determining the time when to stop searching and start executing so as to
provide a near optimal response time. This problem is simplified for small
search spaces (where it may be possible to search the entire search space
quickly) or when it is possible to recognize the best plan (so that the search
can be stopped as soon as the best plan is found). In general it may not be
possible to characterize apriori the best plan and even if it is possible to
characterize it for recognition, reaching it in many search traversal algorithms
may require prohibitively large time.
The basic idea of NORA is simple and intuitive. As mentioned earlier, the
search space is modeled as a set of nodes with inter-connecting arcs. Eacn
node represents one possible plan (or set of actions) and has an associated
execution cost. A search traversal algorithm specifies the order of traversal of
these nodes and there are costs associated with this traversal. These costs
include the cost of actually moving from node to node and any associated
computation costs at each node (for example, to estimate the execution cost of
the plan at the node). The chosen search traversal algorithm goes from node
to node in some order to find the node with the best plan (i.e., the least
action time or equivalently least execution time ). At each node, NORA keeps
track of two metrics:
13
[1] Planning cost so far: this represents all associated costs of planning so
far and includes all cost of traversing the search space. This metric isan umbrella that includes costs of all activities associated with al,
different planning methods used.
[2] Least execution cost so far: this is the least execution cost so far, or
equivalently the best plan found till then.
The search space traversal is terminated whenever the planning cost exceeds
some fraction of the best execution cost found so far, i.e., when the following
condition is satisfied:
least executioncost so far planning cost so far > 6
where (3 defines the fraction. As shown below, this simple heuristic allows the
achievement of a near-optimal response time (when the search is terminated
with the above condition satisfied) satisfying the following bound: iResponse time (optimal) max (1 + 6, 1 +)
This is for the general case, in which we are unable to recognize apriori the
best plan. In cases, where it is possible to recognize the best plan, tighter
bounds can be obtained. Evidently, the choice of 13 is crucial and heuristics
for the choice of an appropriate 13 are also described below. It should be
noted that the bound holds for any given search space traversal ordering
(e.g. A*).
5.2 Correctness of NORA
Response time (obtained whenstopped)
Some notation used in the correctness proof below is introduced in table 1.
14
As mentioned earlier, the search space is structured as a set ot
interconnected nodes Q i , each node representing a possible plan.
Table 1 about here
Correctness Theorem for NORA: If the following search terminating criterion
is used,_,„ ;rki) _ 6
the following upper bound on RT(i) is obtained:RT(i) max(1+13,14)RT(opt)
Proof: Supposing the search terminates after examining Q i . There are thefollowing two cases:
[a] Clopt SP(i), i.e., Qopt has been visited.
[b] Qopt has not been visited.
CASE [a]RT(i) r(i)+t(i)
RT(opt) r(opt) +t(opt)
t4.t.(04. 6
r(opt)+t(opt)since r(i) = ICU ,
B
A positive quantity, ö is added to the numerator to make the equality hold. Itsatisfies the condition, 0 5 5 5. Opt cost in current step.
RT(i) < 6 RT(opt) t(opt) since T(opt) 0
15
(g+1)t(opt)+6
t(opt)
because Q e SP(i) => t(i) t(opt)Opt
1 + 16
Because optimization cost in current step t(opt) => (5/t(opt));-- •0. This
assumption is satisfied by many planners, e.g., query optimizers.
Case [b]:
RT(i) 7(1)+t(1) RT(opt) T(opt)+t(opt)
70)+61-(1) t(i) -r(-Da_ ,Tkopt)+t(opt)
< r(i)+67(i) - T(opt)
< (1+6)T(i)- T(i) -
t(opt)a0
1 + 6
Since Q opt SP(i) => r(opt) r(i). Combining the results of the cases above,
we have:
RIM < max ( 1 +6,1+8)RT( opt)
The bound provided by the above theorem is loose, but the termination
criterion usually performs much better In practice. This was observed during
1 6
experimental validation, the details of which are discussed in the next section.
Corollary 1: If the planning is the compile-and-store type, and is expected tobe executed 1.1 times, the above bound changes to:
RT(i) < max( 1+134, 1+k)RT(opt)
Proof: Identical to the above theorem. (It should be noted that RT(i) = r(i)+mt(i) and represents an integrated cost.)
I5.3 Heuristics For Choice Of @
The choice of 13 depends on the size of the search space, and the timeavailable for planning. For a small search space, Case [a] would be expected tooccur when the search terminates, i.e. Qopt 6- SP(i). ThusSmall space z> Case [a]=> Choose large 13 for tight bound=> Less planning is good for response time
For a large search space, Case [b] would be expected to occur when thesearch terminates, i.e., Qo pt * SP(i). Thus,Large space --z> Case [b]=> Choose small 13 for tight bound=> More planning is good for response time
Rule of Thumb for Choosing 13: The above analysis shows that the thumb ruleis to coose a large 13 if the search space is small and vice versa. It is not acertain rule because of the approximate implications (2:>) shown above.
5.4 General Comments
NORA provides a simple and intuitive, heuristic algorithm for determining thetradeoff between planning and execution times in real time planning andsearch. Two points concerning NORA should be emphasized:
17
[1] The validity of NORA is independent of the search procedure chosen. Thus
any order of traversal specified by any search procedure can be usec
to traverse the search space, and as long as the termination criterion of
NORA is satisfied, the performance bound of NORA shall be satisfied by
the response time obtained.
[2] NORA does not require any apriori domain knowledge. The only metric
needs are some estimates of the execution times of different solutions
and a clock to keep track of the planning time used up so far. It aiso
does not require the goal to be specified apriori .
6 Real-Time Aoolications of NORA
This section demonstrates the application of NORA in two sample domains.:
query optimization (planning) and the A* heuristic search algorithm.
6.1 Semantic Query Optimization
NORA was implemented for optimizing the response time of the query optimizer
for data-bases. Most databases are on-line and users access desirecinformation via a query, represented in a suitable query language. Queries
involve operations like join, select, and project, and can be answered by
executing several different execution-plans each with different execution-times.
The query optimizer generates and examines many execution plans to choose
the one with the least execution-time. The planning time increases
exponentially, as the optimizer expands its search of possible execution plans.
Conventional query optimization (planning) is based on syntacticrearrangements [27] , query decomposition [28], and optimal usage of indices,
join algorithms and database statistics [29]. An orthogonal optimization
technique utilizing application specific knowledge, is Semantic Query
Optimization (SQO) [26]. SQO uses semantic information about the database
instance, eg. semantic integrity constraints and functional dependencies, for
optimization. The original query Q0, is transformed into syntactically different,
but semantically equivalent queries Q1 , Q2 , ..., Qn, with the hope that one of
the semantically equivalent queries may possibly yield a more efficient
18
execution plan than the original query.
Conventional query optimizers select the optimal query execution plan by
searching through QP(0), where QP(0) is the space of query plans
corresponding to the original user query, Q0. In contrast, semantic query
optimizers search the much larger space of execution plans, Q tot,
QP tot = QP( 0 ) QP(1) +....+QP(n)
where QP(i) is the space of query execution plans corresponding to query Qi
and "+" represents the union of the various query plan spaces, QP(i).
The search space can be conceptually modeled in two levels:
[1] Level 1: The space of semantically equivalent but syntactically different
queries, Q0 , Q 1 , ..., Qn.
[2] Level 2: The spaces QP(0), QP(1), QP(n), which are the spaces of
possible execution plans for the queries Q 0, Q 1 , ..., Qn , respectively.
This has been illustrated in figure 5. The solid lines represents the edges at
level_1 search space, and the dotted triangles together constitute the level_2
search space.
Figure 5 about here
As evident, semantic optimization increases the search space of possible plans
by an order of magnitude. This causes the optimization cost to become
comparable to the cost of query execution. The tradeoff between optimization
time and the cost of query execution becomes a major issue In optimizing the
total cost of query processing.
A semantic query optimizer has to search at level_1 of figure 5 to find the
query Q i that leads to the least execution cost. While traversing the space of
semantically equivalent queries, it is not possible to apriori specify the querywith the least execution cost. Thus in the experimental validation, a best first
19
search was used to order the traversal at level_1. The cost of planning hereconsists of the time required to move from one node to the other (i.e., to get
a new semantically equivalent query) and the computational cost at each node
to estimate the execution cost at that node. The execution cost associated with
each node consists of the estimated execution time of the best query executionplan for the query corresponding to that node (obtained by running a
conventional query optimizer on the query associated with that node).
6.2 Experimental Validation
NORA was tested for semantic query optimization for a shipping database of
six relations. The database schema, the relation sizes, semantic integrity
constraints and the various indexes available are reported in detail in [34, 35].
The first step in the experiment was to generate the entire space of
semantically equivalent queries. This was done manually by applying rules, i.e.semantic integrity constraints, giving rise to queries Q 1 through Q15, which
were all semantically equivalent to an input user query Q .0
Figure 6 about here
Figure 6 shows the entire generated search space of semantically equivalent
queries. In figure 6, each node (box) represents a semantically equivalent but
syntactically distinct query. The numbers on top of each node are the query
numbers, i.e., alternative semantically equivalent queries. The arrow points to
the input user query Q0 . Each query is a conjunction of the clauses Ci's.
Rules (semantic/Integrity constraints) R i 's perform the transformationsbetween queries by adding or deleting clauses. Each of the queries Q0
through Q15 was optimized by the R* [30] optimizer to obtain the best
estimated execution costs associated with each query (shown as numbers
within the boxes in figure 6). The NORA algorithm was hand-simulated on the
graph shown in figure 6. NORA terminates the search when the following
condition becomes true.
r(i) > t 8i
20
Table 2 about here
The stopping rule was examined for the values 13 = 2, 1, and The results
are presented in Table 2. The result can be verified for all the three values
of 13. The validation for 13=1 is illustrated below. The search stops after
iteration 3, since the stopping criterion of NORA evaluates to
(163)(1) + (5)(5) > 127.0 => 188 > 127.0
which is true. The optimization cost and best execution cost estimate are
r(i) = 188 ms; t(i) = 127.0 ms
The bound of the Correctness Theorem is satisfied since
7(1)+t(i) 315,r(opt)+t(opt)-224-2
It can be seen that stopping rule is quite effective. Especially notable is the
fact that even though it was attempted to minimize a weighted sum of t(i) anc
t(i), and not t(i) alone, the value of t(i) obtained is actually quite close to the
minimum.
6.3 Ax Heuristic Search Algorithm
A* [14] is probably the best known heuristic search algorithm for finding a
solution path from an initial start node to a final goal node. A* essentially
does a best-first search of the search space, where the merit of a node, f(n),
is the sum of the actual cost, g(n), of reaching that node, n, from the start
node, S, and the estimated cost, h(n), of reaching a goal node, G, from that
node. A* has the property that it will always find an optimal solution to a
problem if the heuristic function, h(n), is admissible, i.e., it never
overestimates the actual cost of reaching a goal node. Further if the heuristic
function h, satisfies the monotone restriction [14] (i.e., the estimate of the
optimal cost to a goal from node n i not be more than the cost of the arc from
node n i to ni plus the estimate of the optimal cost from ni to a goal), then it
can be proved that A* already has found an optimal path to any node that it
selects for expansion, i.e., it never expands any nodes not on the optimal path
from the start node to a goal node.
21
A* ignores execution time and essentially concentrates on searching the search
space for an optimal solution path. As described in section 4, even the real-
time variations on A* proposed by Korf ignore execution time. This is not
desirable always as explained earlier. To apply NORA or any related formalism
to A*, the notion of execution time has to be introduced. It should be notea
that in A*, the implicit assumption of being able to apriori recognize the goa'
node is made (necessary for the evaluation of the heuristic function h). This
is in contrast to the more general framework of NORA and the semantic query
optimization example described earlier where it is assumed that it is not
possible to apriori recognize the goal node.
Associated with any node n, in the search space, let ge(n) represent the
execution time for executing the partial solution (corresponding to the path
from the start node to node n) found so far and let he(n) represent the
heuristic estimate of the cost of execution from the partial solution to a goa!
node (corresponding to a path from node n to a goal node). Note that ge(n)
and h e(n) are distinct from the conventional g(n) and h(n) functions of A*.
While g(n) estimates the cost of expanding the search tree from the start node
to the node n and thus contributes to the planning costs for finding a
solution path from the initial state to the partial solution state specified by
node n, ge(n) estimates the cost of actually executing that partial solution
path in the real world. Similarly, while h(n) estimates the planning cost of
further expanding the search tree from node n to a goal node, he(n)
estimates the actual cost of execution while trying to reach a goal from the
partial solution state specified by node n. These functions can be computed as
it is possible to apriori precisely characterize the goal state.
While trying to optimize response times in real time planning and search using
A*, it is possible to decide to terminate the search (based on a chosen search
stopping criterion, e.g., as specified by NORA) at some node i, before reaching
a goal node, G. Assume that stopping at some node i, (i * goal node), forces a
penalty for execution in real time (caused by the distance from the partial
solution state, node i, to a goal node), and let it be represented by p(i). It is
assumed that p(i) is proportional to the distance of the node i from a goal
node and that it monotonically decreases as i approaches a goal node.
22
6.3.1 Aoolyino NORA to A*
Assume that the A* heuristic evaluation function, h, satisfies the monotone
restriction and is admissible. Under these conditions, A* never expands ar,.,
nodes other than those that lie on the optimal solution path and thus at any
node, (i goal node), on the optimal solution path, the total planning cos:
so far is given by g(i) and the best estimate of the execution cost is given ty
the sum of g e(i) and p(i). Let G represent the goal node, and thus the
planning cost for reaching the goal node is g(G) and the execution cost is
given by the sum of g e(G). The stopping criterion as specified by NORA woulc
be:
stop when:
9(i) e (1)+P(1)
Writing the ratio of the response time, RT(i), when stopping at node i over the
response time, RT(G), when stopping at goal node, G, we have:
RT(i) 9(i)+9.(i)+P(i) RT(G) g(G) + 9.(G)
g(i)+9,(i)+p(i)- g(i)+h(i)+9.(G)
Note that g(G) g(i) + h(i) as h is assumed to be admissible. Since h(i) is
positive, it can be ignored from the denominator and on substituting the
stopping condition, we have:
(1+6)g(i)- 9(1)+9.(G)
Dividing both numerator and denominator by g(i):
1+6 g '(G)
1+ 9(i )
23
as both ge(G) and g(i) are positive. Thus the bound of NORA is satisfied. Note
that as 13 is positive, 1+13, is greater than 1.
D.3.2 An Alternative Heuristic Far A*
Having apriori precise characterization of the goal allows the formulation of
another stopping criterion for A*, which uses more domain specific knowledge
(and thus is less general than NORA), but leads to a crisper bound. Assume
that the function h e(i) never over-estimates the real execution time and that
the heuristic evaluation function, h, is admissible and satisfies the monotone
restriction.
Intuitively, it is clear that a useful stopping criterion is to stop when the
following condition holds:
h(i) + he(i) ap(i)
where a is some positive constant decided by the user. Here the LHS is the
sum of the estimated planning and execution costs for reaching a goal state
from the current node, i, and the RHS give the penalty incurred for stopping
at node i and not continuing onto reach a goal node. It is obviously better to
stop at node (i) and incur the corresponding penalty, p(i), rather than spend
the effort, h(i) + h e(i), towards planning and execution from that node i
(towards the goal). Under these conditions:
RT(i) 9(i)+9.(1)+p(i) RT(G) g(G) + g,(G)
Note that as both functions, h(.) and he(.) never overestimate values, the
following substitution can be made in the denominator:
g(i )+g , (i )+p(i)
g(i)+h(i)+g.(i) +h,,(1)
Substituting the stopping criterion, we have:
24
g(i)+Mi)+p(i)- 9(i)+9.(i)+0P(i)
where
p(i)x - g(1)+g.(1)
Here X can be interpreted as the ratio of the penalty p(i), for stopping
prematurely at node i, and the total response time, g(i) + g e(i). Note that the
bound obtained in this case is less than 1, when o is positive and greater
than 1. In contrast, the bound obtained using NORA was larger than 1.
7 Planning & Search Under Deadlines
Section 4 reviewed some work by other researchers on planning under
deadlines. Korf [6, 7] proposed RTA*, a real time modification of A*, to perform
real time heuristic search under deadlines. Korf's algorithms essentially spends
all available time searching/planning and ignore execution costs. To introduce
execution costs into the scenario, RTA* would have to be modified to keep
track of a dynamic deadline , which is in effect the real deadline less the best
execution cost found so far, i.e.,
dynamic deadline = real deadline - least execution cost
The real deadline is set by the external world, and the dynamic deadline is a
virtual deadline of which the system keeps track. The search/planning is to
be terminated as soon as the dynamic deadline is reached. The dynamic
deadline can vary in run time depending upon the variations of the least
execution costs. This is a simple extension to the work of Korf and thus most
of Korf's analysis shall carry over to this modified RTA* also.
Bonissone's [31] approach to planning/search under deadlines has been to
25
precompile the proof tree/search space along with detailed estimates of thecosts of traversing any one path. In run time, depending upon the deadline,
some or all different paths in the proof tree/search space are explored. Some
heuristics are also provided to decide which path to explore next. This
approach is well-suited for well structured domains in which such knowledge
is available apriori.
8 Plannina & Search Using Diminishing Marginal Returns
Heuristic search strategies like best-first usually have diminishing marginal
returns as the search progresses. This can be illustrated by the example
discussed
in section 6.2. Database access paths are tailored to suit the
frequently occurring queries, and semantic transformation based on integrity
constraints can not keep improving the execution cost for a long time. The
optimal query Q min
Q min = Q 1 such that C E (i) 5 CE (j), j=0,1,..,n
has a finite positive cost, and as it is approached during the search, the
chances of substantial improvement in query execution cost keep diminishing.
At the same time the optimization cost remains approximately the same for
every step of searching in the space of queries. Thus, net improvement in
query execution cost keeps diminishing.
A heuristic based on diminishing marginal returns can be used to consider the
net benefit obtained from the last step and decide appropriately the tradeoff
between planning and execution costs. The basic Idea, simply put, is to stop
searching when the improvement in execution costs is dominated by the
planning cost from the last step (node). This termination criterion leads to the
optimal solution only if the law of diminishing marginal utility holds. However,
even if this is not true, the criterion Is still useful for small search spaces.
The searching process may be terminated at a local minima. This can be
partially overcome by using probabilistic search guiding strategies, eg.
simulated annealing [32, 33]. Note however that this heuristic cannot guarantee
either a near optimal response time or response within a deadline.
This can be illustrated with the example of sections 6.1 and 6.2. Search stops
when the optimization cost in last step dominates the improvement in query
26
execution cost, i.e., (using the data of table 2):
Reduction in t(i) < Optimization Cost
At step 1,
Reduction in t(i) in step 1 = inf - 605.7 = m ms
Optimization Cost in step 1 = (10)(1) = 10 ms
At step 2,
Reduction in t(i) in step 2 = 605.7 - 127.8 = 477.9 ms
Optimization Cost in step 2 = (66)(1) + (3)(5) = 81 ms
At step 3,
Reduction in t(i) in step 3 = 127.8 - 127.0 = 0.8 ms
Optimization Cost in step 3 = (87)(1) + (2)(5) = 97 ms
Thus, the search stops after step 3, having incurred a total optimization cos:
of 188 ms, and having generated a query execution plan with an estimated
cost of 127.0 ms. The above analysis shows that this heuristic based or
diminishing marginal returns can be useful since even for an interactive
query (i.e. execute only once) the savings obtained are
605.7 - (188.0 + 127.0) = 605.7 - 315.0 = 290.7 ms .
If the query is to be executed many times the savings are even greater.
9 Summary
This paper has tried to formalize ideas relating to approximate reasoning about
meta-level temporal constraints. The results obtained in this paper are
important because they are the among the first in this area within artificial
intelligence. As mentioned earlier, real time problems are gaining importance in
artificial intelligence. The algorithm developed in this paper will enable
decisions to be made regarding the crucial tradeoff of relative durations of
planning and execution times. Issues from this paper worth emphasizing are:
■ Reasoning about the relative durations of planning and execution times is an
important problem in real time problem solving.
■ Real time reasoning and problem solving can be conceptually modeled as a
search among possible solutions.
27
■ Heuristics are best for determining the tradeoff between planning and
execution times in general search spaces, with weak domain knowledge.
■ NORA uses the following intuitively, simple heuristic as the termination
condition: stop planning when planning costs exceed a fixed fraction of
the best execution cost found so far .
■
NORA is valid for any search algorithm and requires no apriori domain
knowledge.
■ NORA
has been applied in the sample domains of semantic query optimization
and the A* heuristic search algorithm.
■ For planning under deadlines, Korf's algorithms can be modified with a
dynamic deadline .
■ The
concept of diminishing marginal returns can also be used as a stopping
heuristic in real time reasoning.
References
1. R.P. Bonaso, "What AI Can Do for Battle Management: A Report of the AAAIWorkshop on AI Applications to Battle Management," AI Magazine, vol. 9,no. 3, Fall 1988.
2. H.P. Newquist III, "American Express and AI: Don't Leave Home WithoutThem", AI EApert, vol. 2, no. 4, Miller Freeman Publications,500Howard St., San Francisco, CA 94105., April 1988.
3. D. Galperin, "On the optimality of A*," Artificial Intelligence, vol.8, no. 1,pp. 69-76, 1977.
4. R.E. Korf, "Depth-First Iterative Deepening: An Optimal Admissible TreeSearch, " Artificial Intelligence, vol. 27, pp. 97-109, North-Holland,1985.
5. T.J. Laffey and P.A. Cox, "Real-Time Knowledge Based Systems," AIMagazine, vol. 9, no. 1, Spring 1988.
6. R.E. Korf, "Real-Time Heuristic Search: First Results," Proc. AAAIConference, 1987.
7. R.E. Korf, "Real-Time Heuristic Search: New Results," Proc. AAAIConference, 1988.
28
8. M.R. Garey and D.S. Johnson, "Computers and intractability: A guide to thetheory of NP-completeness", W.H. Freeman, New York, 1979.
9. C.A. O'Reilly and A.S. Cromarty, "Fast is not Real-Time in DesigningEffective Real-Time AI Systems," Applications of Artificial IntelligenceII, International Society of Optical Engineers, Bellingham, Washington,1985.
10. J. Marsh and J. Greenwood, "Real Time AI: Software Architecture Issues,Proc. IEEE Nat'l Aerospace and Electronics Conference, WashingtonD.C., 1986.
11. D.E. Wilkins, "Practical planning", Morgan Kaufmann Publishers,PaloAlto, CA, 1988.
12. D. Chapman, "Planning for conjunctive goals," Artificial Intelligence, vol.32, pp. 333-378, July 1987.
13. M.P. Georgeff, "Planning," Annual review of computer science, vol. 2, pp.359-400, 1987.
14. N.J. Nilsson, "Principles of Artificial Intelligence", Tioga, Palo Atto, CA,1980.
15. R.E. Korf, "Planning as search: a quantitative approach," ArtificialIntelligence, 1987.
16. D.E. Wilkins, "Domain-independent planning: representation and plangeneration," Artificial Intelligence, vol. 22, pp. 269-301, April 1984.
17. S. Russell and E. Wefald, "Decision Theoretic Control of Reasoning: Genera,Theory and an Algorithm to Game Playing," Report No. UCB/CSDB8/435,Computer Science Division, U.C. Berkeley, 1988.
18. E.J. Horvitz, G.F. Cooper, and D.E. Heckerman, "Reflection and actionunder scarce resources: theoretical principles and empirical study," InProc. of IJCAI, Detroit, MI, Aug., 1989.
19. E.J. Horvitz, "Reasoning about beliefs and actions under computationalresource constraints", Proc. of the 3rd workshop on uncertainty inartificial intelligence, AAAI, Seattle, WA, July, 1987.
20. E.J. Horvitz, "Reasoning under varying and uncertainresourceconstraints," In Proc. of AAAI, pp. 111-116, Minneapolis, MN,Aug.,1986.
21. I.J. Good, "Rational decisions," J.R. Statistics Soc. B, vol. 14, pp./07-114,1952.
22. J.E. Matheson, "The value of analysis and computation," IEEE Transactionson Systems, Science and Cybernetics, vol. 4, pp. 211-219, 1968.
23. S.R. Watson and R.V. Brown, "The value of decision analys=s," J.R.
29
Statistical Soc. A., vol. 141, pp. 69-78, 1978.
24. D.V. Lindley, "Reconciliation of decision analyses," Operations Research,vol. 34 (2), pp. 289-295, 1986.
25. P.P. Bonissone and P. Halverson, "Time constrained reasoning withuncertainty," Workshop on Real Time Problems in AI, IJCAI, Detroit,MI, 1989.
26. J.J. King, "QUIST: A system for semantic query optimization in relationaldatabases," Proc. 7th VLDB Conf., 1981.
27. Smith and Chang, "Optimizing the performance of a relational algebradatabase interface," CACM, vol. 18:10, 1975.
28. E.
Wong and K. Youseffi, "Decomposition - A strategy for quer;optimization," ACM TODS, Sept. 1976.
29. P. Selinger, "Access path selection in a Relational Database ManagementSystem," RJ 2883, IBM San Jose, 1979.
30. L.F. Mackert and G.M. Lohman, "R* Optimizer Validation and PerformanceEvaluation for Local Queries," Proc. ACM-SIGMOD, pp. 84-95, ACM,1986.
31. P. Bonissone, Private communication, 1989.
32. Y.E. Ioannidis and Wong, E., "Query Optimization by Simulated &Annealing," Proc. SIGMOD, 1987.
33. S. Kirkpatrick, et. al., "Optimization by Simulated Annealing, Science,"pp. 671-680, May 1983.
34. S.
Dutta, "Approximate reasoning with temporal and spatial concepts", Ph.D.thesis, Dept. of Computer Science, U.C. Berkeley, 1990.
35. J. Srivastava, S. Shekhar & S. Dutta, "A formal model of trade-off betweenoptimization and execution costs in semantic query optimization", In theProc. of the Int'l Conf. on Very Large Databases, LA, Aug. 1988.
Figure 1: Three phases of real time problem solving
Data Planning & Execution of
Collection
Reasoning I Solution
• •To TI T2 T3
Time
Figure 2: Temporal decomposition of real time problem solving
Deadlines
z
Real Tune Constraints
on Response Time
Optimal ResponseTimes
CompleteSolution
Rana!
Solution
Figure 3: Classification of Real Time Systems
TIME
EXECUTION............ TIME
OPTIMIZATION EFFORT
Figure 4: Planning, Execution and Response Times
Fig. 5 Conventional Vs. Semantic Query Optimization.
4
(CI ,C3,C4 ,C5127.8
>
7.
R2
_ _ R6_ _
R2
1 'R7 - -
R7_ .1 - •
- - / R116
2
,A1
Fig. 6 Space of Semantically Equivalent Queries.
Symbol Meaning
SP (i)1
space of nodes already exploredupto and including Q,
CE v)
cost of best executionplan for node Qi
-c(i )total planning cost upto
node Q,
t (i)cost of best plan
so far, i.e min CE(j)Q, c SP (1)
RT (i)response time if search
terminates at Q,,i.e. TO )+ t (i)
RT (opt)
theoretically minimum response time
possible for the given search algorithm
RT (opt) .-- min RT (i)osi se,
t(opt), t (opt) components of RT (opt)
QopI node corresponding to RT (opt)
Table I Notation
Iteration t(i) t(i) TO )+t (i ) Stopping(p) RT (i )/ RT (opt)
1 10 605.7 615.7
2 97 127.8 224.8 p =2 I
3 188 127.0 315.0 p =I 1.5
4 359 127.0 486.01
fiz----i. 2.5
5 647 127.0 774.0
Table 2 Performance of NORA
INSEAD 908E180 PAPERS SERIES
1986
86/01
Arnoud DE MEYER
86/02 Philippe A. NAERTMarcel VEVERBERCBand Guido VERSV1JVEL
86/03 Michael BRIMM
86/04 Spyros MARRIDAXIS•ndMichilealSON
86/05 Charles A. VTFLOSZ
86/06 Francesco CIAVANI,Jeff R. SHEN andCharles A. VYPLOSZ
86/07 Douglas L. MacLACHLANend Spyros MAKRIDAXIS
86/08 Jos4 de la TORRE andDavid NECKAR
86/09 Philippe C. RASPESLACII
86/10 R. MOENART,Arnoud DE METER,J. BARBE andD. DESCROOLMEESTER.
86/11 Philippe A. NAEATand Alain BULTEZ
86/12 Roger BiTANCOURTend David GAlITSCIII
86/13 S.P. ANDERSONand Damien J. NEVEN
86/14 Charles VALDNAN
' The R L 0/Production loterface.
*Subjective estimation In Integratingcommunication budget and allocationdecisions: a case study', January 1986.
• Sponsorship and the diffusion oforganisational innovations a preliminary vlev'.
' Confldeoce Intervals: an empiricaliovestigetion for the series in the K-
Coomaltiono
'A note on the reduction of the vorkveek',July 1985.
'The real exchange rate and the fiscalaspect. of • natural resource discovery•,Revised version' February 1986.
'Judgmental biases In sales forecasting•,February 1986.
• forecasting political risks forinternational operations • , Second Draft:March 3, 1986.
"Conceptualising the strategic process indiversified firms, the role and nature of thecorporate influence ItITOC412, • , February 1986.
'Analysing the Issues concerningtechnological do-maturity•.
' Prom 'Lydlametry* to 'Pinkhamication't■isspecitying advertising dynamics rarelyaffects profitability'.
'The economics of retail firms'. RevisedApril 1906.
'Spatial competition i la Cournot'.
'Comparalson Internationale des merges brutesdu commerce', June 1985.
'Les prise. des offres publiques, I. noted'intornatioo et le &trait des transterts decontrOle cit. socIdtts'.
• Stratesic capability transfer in scoultoltlunIntegratioe, May 1986.
'Towards en operational definition ofservices', 1986.
'Nostradamust a knovledge-based forecastingadvisor'.
'The pricing of equity on the London stockexchange: seasonality and site premium',June 1986.
'Rlsk-premla seasonality in U.S. end Europeanequity markets', February 1986.
'Seasonality In the risk-return relationshipssome international evidence', July 1986.
'An exploratory study on the integration ofinforsmtIon systems In manufaeturing•,July 1986.
'A methodology for specification andaggregation In product concept testing'.July 1986.
' Protection'. August 1986.
' The economic consequences of the FrancPoint:kW, September 1906.
'Negative risk-return relationships Inbusiness tttttt rya paradox or trulsof'.October 1986.
'Interpreting organizational tests.
"49,y follov'the leader?•.
'The succession CAA!, the real story.
' flexibility: the neat competitive battle',October 1986.
"flexibility: the nest competitive bottle',Revised Version! March 1987
86/16 8. Espen ECK./10 andHervig M. LANCOHR
86/17 David ft. JEMISON
86/18 Janes ICBMAnd V. MALLEKET
86/19 Rob R. trC1T2
86/20 Albert CORNAT,Gabriel UMAVINIand Pierre A. M/CIIEL
86/21 Albert CORMAT,Gabriel A. ilmtissdiNiand Pierre A. MICHEL
86/22 Albert CORnAT,Gabriel A. RAVAVINIand Pierre A. MICSEL
86/23 Arnoud DE METER
86/24 David CAUTSCRIand Vithala R. RAO
86/25 R. Peter CRAYAnd Ingo VALTER
86/26 Barry C1CRENGREENand Charles vynosz
86/27 Karel COOLand Ingemar DIERICK1
86/28 Manfred KETS DEVRIES and Danny MILLER
86/29 Manfred KLTS DC VRIES
86/30 Manfred KEYS DE VRIES
86/31 Arnoud DE METER
86/31 Arnoud DE METER,Jinichiro NAKANE,Jeffrey C. MILLERAnd )(sera FERDOVS
86/15 MIhkel EMU andArnoud DE METER
• lov the managerial attitudes of firma vithFMS differ from other manufacturing firms,surve y results'. June 1986.
86/32 Karel COOLA nd DAn scnr.nort.
Performance differences song strategic groupO<Ob4f7 • . October 1906.
87/06 Arun K. JAIN, 'Customer loyalty es a construct in theChristian PINSON and marketing of banking services'. July 1986.Naresh K. NALNOTRA
86/33 Ernst BALTENSFERCERand Jean °ERMINE
86/34 Philippe MASPESLAGM
and David JEMISON
06/35 Jean DERWIN!
86/36 Albert CORRAT and
Gabriel MAVAVINI
86/37 David GAVISCnI and
Roger BETANCOURT
86/38 Cab r lel HAVAVINI
86/39 Gabriel RAVAVINIPierre MICRELand Albert CORRAT
86140 Charles 'rerun%
86/41 Kase& FERDOVS
and Uickham SKINNER
86/42 Kama FERDOVSand Per LINDBERG
86/43 Damien NEVER
86/44 In g DIERICKX
Carmen MATUTESand Damien NEVEM
1987
87/01 Manfred KETS DE VRIES
87/02 Claude VIALLET
87/03 David CAUTSCRIand Vithal• RAO
87/04 Sumantra CHOSHAL and
Christopher BARTLEY,'
87/05 Arnoud DE METERand fast. • FERDOVS
'The role of public policy In insuringfinancial stability: • crass-country,cooperative perspective', August 1986, Revised
November 1986.
•Acquisitions: myths and reality',
July 1986.
'Measuring the market value of • bank, •
primer', November 1986.
'Seasonality in the tisk-return relationships
some international evidence', July 1986.
'The evolution of retelling: • suggested
•conooic interpretation•.
'Financial innovation and recent developmentsIn the French capital markets', Updated:
September 1986.
'The pricing of common stocks on the Brusselsstock eachangel • re-elanInation of theevidence • , November 1986.
*Capital flows llberalizstlon and the EliS, •french perspective', December 1986.
°Manufacturing in • new perspective',
July 1986.
erns as indicator of aanufacturing strategy',
December 1966.
'On the 41[111t•GO of equilibrium in hottalling'•model', November 1986.
•Value added tax and competition',December 1986.
'Prisoners of leadership'.
'An empirical Investigation of international
asset pricing', November 1986.
'A methodology for specification and
aggregation in product concept testing',
Revised Version: January 1987.
'Organising for innovations: case of the
multinational corporation', February 1987.
'Managerial focal points 1n manufacturingstrategy', February 1987.
'Nulty pricing and stock market anomalies',
February 1987.
'Leaders who can't manage', February 1987.
'Entrepreneurial activities of European RDAs',
March 1987.
'A cultural view of organizational change',
March 1987
'Forecasting and loss functions', March 1987.
'The Janus Dead: learning from the superiorand subordinate feces of the manager's job',
April 1987.
'Multinational eorporatiors as differentiatednetwork'', April 1907.
'Product Standards and Competitive Strategy: An
Analysis of the Principles', May 1987.
•METAYORICASTING: Vays of improvingForecasting. Accuracy and Usefulness',
May 1987.
'Takeover attempts: vhat does the language tell
us7, June 1987.
•Managers' cognitive maps for upvard anddownward relationships', June 1987.
"Patents and the Ruropean biotechnology lag: •study of large European pharmaceutical fires',
June 1987.
'Vhy the EMS? Dynamic games and the equilibrium
policy regime. Hay 1987.
'A nev approach to statistical forecasting',
June 1987.
'Strategy formulation: the impact of national
culture', Revised: July 1987.
'Conflicting ideologies: structural and
motivational consequences', August Iwo.
'The demand for retail product, and thehousehold production 'model: nev views on
complemientarity and substitutability'.
87/07 Rolf BANE andGabriel NAVAVINI
87/08 Manfred KETS DE VRIES
87/09 Lister VICKERY,
Mark PILKINCTON
and Paul READ
87/10 Andre LAURENT
87/11 Robert FILDES and
ShYrosNAKEIDAXIS
87/12 Fernando BARTOLOMEand Andre LAURENT
87/13 Sumantra CHOSHALand Nitin NOIIRIA
07/14 Landis GABEL
87/15 Spyros KAKRIDAKIS
87/16 Susan SCHNEIDER
and Roger DUNBAR
87/17 Andre LAURENT and
Fernando BARTOLOME
87/10 Reinhard ANGEL/4AR and
Christoph LIEBSCHER
87/19 David BEGG and
Charles VTPLOSt
07/20 Spyros MAKAIDAXIS
87/21 Susan SCHNEIDER
87/22 Susan SCHNEIDER
87/23 Roger BETANCOURTDavid GAUTSCIII
'Spatial competition and the Core', August
1987. 1988
88/10 Bernard SINCLAIR-
DESCAGNé
87/40 88/11 Bernard SINCLAIR-DESCACNt
Carmen NATIVES andPierre RECIBEA0
'The robust t some standard auction game
torus*, February 1988.
'ifheo stationary strategies are equilibriumbidding strategy: The single-crossing
property • , February 1988.
87/39 Manfred KM'S DE VRIES 'The dark side of CEO succession", November
1987
'Product compatibility and the scope of entry•,November 1987
•The internal and external careers: atheoretical and cross-cultural perspective',Spring 1987.
•The robustness of KDS configurations In theface of incomplete data • , March 1987, Revised,July 1987.
"Deanna cosplementarities, household productionand retail assortments', July 1907.
'Is there • capital shortage In Europe?',August 1987.
'Controlling the Interest-rate risk of bonds:an introduction to duration analysts andImmunisation strategies', Sep t ember 1987.
•Interpreting strategic behavior: basicassumptions themes in organizations', September1987
'On the optimality of central places*,September 1987.
*German, French and British manufacturingstrategies less different than one thinks',September 1987.
*A process framework for analyzing cooperationbetveen firms • , September 1987.
•European manufacturers: the dangers ofcomplacency. Insights from the 1907 Europeanmanufacturing futures survey, October 1987.
*Privatisation: its motives and likely
consequences', October 1907.
•Strategy formulation: the impact of nationalculture • , October 1987.
87/41 cavriel HAVAVIN1 andClaude VIALLET
87/42 Nolen NEVER andJacques-P. THISSE
87/43 Jean GABSZEV1CZ andJacques- F.
87/44 Jonathan HAMILTON,Jacques-P. THISSEand Anita VESKANP
87/45 Karel COOL,David JEMISON andIngenar DTERICK1
87/46 Ingenar DIERICKIand Karel COOL
88/01 Michael LAVRENCE andSpyros KAKAIDAKIS
88/02 Spyros MAKRIDAKIS
88/03 James TEBOUL
88/04 Susan SCHNEIDER
88105 Charles VTPLOSZ
88/06 Reinhard ANSELMAR
88/07 Ingenar DIERICKX
and Karel COOL
88/08 Reinhard ANGELMAA
and Susan SCHNEIDER
88/09 Bernard SINCLAIR-DESGACNf
'Seasonality, sire premium and the relationshipbetween the risk end the return of Frenchcommon stocks', November 1987
'Combining horizontal and verticaldifferentiation: the principle of mma-mindifferentiation • , December 1987
'Location', December 1907
'Spatial discrimination: Bertrand vs. Cournotin • model of location choice • , December 1907
"Business strategy, ameket structure and risk-return relationships, • causal interpretation',December 1987.
'Asset stock accumulation and sustainabilltyof coapetitive advantage', December 1187.
'factors affecting judgemental forecasts and
confidence intervals', January 1908.
"Predicting recessions and other turningpoints", January 1988.
'De-Industrialize service for quality • , January
1988.
'National vs. corporate culture: Implicationsfor human resource management', January 1900.
"The swinging dollar: Is Europe out of step?".
January 1988.
• Les conflits dans les canal,* de dIstributine,
January 1988.
'Competitive advantage: a resource
perspective', January 1980.
'Issues in the study of organizational
cognition*, February 1989.
"Price formation and product design throughbidding', February 1908.
87/24 C.B. DERR andandr4 LAURENT
87/25 A. K. JAN,N. K. MALHOTRA andChristian PINSoN
87/26 Roger BETANCOURTand David GAUTSGHI
87/27 Michael BURDA
07/28 Gabriel HAVAVINI
87/29 Susan SCHNEIDER andPaul SHRIVASTAVA
87/30 Jonathan HAMILTONV. Bentley 'MCLEODand J. 7. THISSE
87/31 Martine OUINEII andJ. P. TIIISSE
87/32 Arnoud DE METER
87/31 Yves DOZ andAmy SHUN
82/34 Kasra FERDOVS andArnoud DE METER
87/37 Landis GABEL
87/38 Susan SCHNEIDER
87/35 P. J. LEDERER and •Competitive location on networks under
J. F. THISSE discriminatory pricing • , September 1987.
87/36 Manfred KETS DE VRIES "Prisoners of leadership', Revised version
October 1987.
88/12 Spyros MAKRIDAKIS
88/13 Manfred KETS DC VRIES
08/14 Alain NOEL
88/15 Anil DEOLALIKAA andLars-Hendrik ROLLER
88/16 Gabriel HAVAVINI
88/17 Michael BURDA
88/18 Michael BUROA
88/19 M.J. LAVRENCE andSpyros MAKAIDAKIS
88/20 Jean DERMINE,Damien NEVEM andJ.F. TIIISSE
88/21 James TEBOUL
88/22 Lars-Hendrik RULER
88/23 Sjur Didrik FLANand Georges ZACCOUR
88/24 B. Espen DCKBO andHervig LANCOHR
88/25 Everette S. GARDNERand Spyros MAKAIDAKIS
88/26 Sjur Didrik FIRMand Georges ZACCOUR
88/27 Murugapps KRISHNANLars-Hendrik ROLLER
88/28 Su•antra GAOSRAL andC. A. BARTLETT
' 8ustness firms and managers In the 21stcentury • , February 1988
'Alexithymis to organisational life: theorganisation tun revisited', February 1988.
' The interpretation of strategies: ■ study ofthe Impact of CEOs on the corporation',March 1988.
"The production of and returns from industrialInnovations an econometric analysis for •developing country • . December 1987.
' Market efficiency and equity pricing,international evidence and implications forglobal I int', March 1908.
' Monopolistic competition, coats of adjustmentand the behavior of guropean employment',September 1987.
'Reflections on Walt Unemployment* inEurope• , November 1987, revised February 1988.
' Individual bias In judgements of confidence',March 1988.
'Portfolio selection by mutual fund,. nnequilibrium andel'. March 1988.
"De-industrialite service (or quality".March 1988 (88/03 Revised).
"Proper Ouadratic Functions vith an Applicationto AT&T', May 1987 (Revised March 1988).
'BquIllbres de Mash-Cournot dans le marchfeuroplen du gal: man gag oil len solutions enboucle ouverte et en feedback coincident•,Mars 1988
' Inforiaation disclosure, 'Deans of payment, andtakeover pees's. Public and Private tenderoffer, in France', July 1985, Sixth revision,Apr11 1988.
' The future of forecasting', April 1988.
' Seal-competitive Cournot equilibrium inmultistage oligopolies • , April 1908.
Matey game vith resalable capacity',April 1988.
'The multinational corporation a, • network:pe-apectives from interorg•nizational theory',w. w toes
80/29 Haccah K. MALROTRA,Christian PINSON andArun K. JAIN
88/30 Catherine C. ECKELand Theo VERMAELEN
88/31 Sumantra GROSMAL andChristopher BARTLE1T
88/32 Kiwi FERDOVS andDavid SACKAIDER
88/33 Mihkel M. TOMBAK
88/34 Mihkel N. TOMBAK
88/35 Mihkel N. TOMBAK
88/36 Vikas TIBREVALA andBruce BUCHANAN
80/37
Nurw trappa KR1SRNANLars-Hendrik ROLLER
00/38 Manfred KETS DE VRIES
88/39 Manfred VETS DE VRIES
88/40 Josef LAXONISROK andTheo VERAAELEN
88/41
Charles VIPLOSZ
88/42 Paul EVANS
88/43 B. SINCLAIR-DESCAGNE
88/44 Essam MAHMOUD andSpyros KAKRIDAXIS
88/45 Robert KORAJCZYKAnd Claude VIALLFT
88/46 Yves DOZ andAmy SHUEN
'Consvwer cognitive coaplex1ty and thedimensionality of multidieensional scalingconfigurations', May 1988.
'The financial fallout from Chernobyl: riskperceptions and regulatory response'. May 1988.
'Creation. adoption, and diffusion ofinnovations by subsidiaries of multinationalcorporations', June 1988.
' International manufacturing: positioningplants for success', June 1988.
' The isportance of flexibility Inmanufacturing', June 1988.
' Flexibility: an important diatenxion inmanufacturing', June 1988.
' A strategic analysis of investment In flexiblemanufacturing systeas • , July 1988.
'A Predictive Test of the N110 Model thatControls for Non.stationarity% June 1988.
' Regulating Price-Liability Competition ToUprose Velfare', July 1988.
'The Motivating Role of Envy : A ForgottenFactor in management, April 88.
' The Leader as Mirror : Clinical Reflections',July 1908.
' Anomalous price behavior around repurchasetender offers', August 1988.
' Assymetry in the EMS, intentional orsystemic'', August 1988.
'Organizational development In thetransnational prise', June 1988.
'Croup decision support systems implementBayesian rationality'. September 1988.
'The state of the art and future directionsin combining forecastn • , September 1988.
' An empirical investigation of internationalasset pricing'. November 1986, revised August1988.
'Prom Intent to outcome: • process fraaevorkfor partnerships', August I980.
88/52 Susan SCHNEIDER andReinhard ANGELMAR
"Cognition and organizational analymiso vho'sminding the store", September 1988. 1989
88/47 Alain BULTEZ,Els CIJSBRECHTS,Philippe NAERT andPiet VANDEN ABEELE
88/48 Michael BURDA
88/49 Nathalie DIERKENS
88/50 Rob VEITZ andArnoud DE MEYER
88/51 Rob VEITZ
' Asymmetric cannibalism betveen substituteitems listed by retailers', September 1988.
"Reflections on 'Veit unemployment' inEurope. II", April 1988 revised September 1988.
"Information asymmetry and equity Issues',September 1988.
' Managing expert systems: from inceptionthrough updating', October 1987.
"Technology, mock, and the organization: thel.pact of expert system', July 1988.
88/63 Fernando NASCIMENTOand Viltried R.VANHONACKER
88/64 Kasra FERDOVS
88/65 Arnoud DE METERand Kasra FERDOVS
88/66 Nathalie DIERKENS
88/67 Paul S. ADLER andKasra FERDOVS
'Strategic pricing of differentiated consumerdurables in • dynamic duopoly: ■ numericalanalysis*. October 1988.
"Charting strategic roles for internationalfactories• , December 1988.
'Quality up, technology dove", October 1988.
"A discussion of exact measures of informationasymmetry: the example of Myers and NONEmodel or the importance of the asset structureof the firm', December 1988.
"The chief technology officer", December 1988.
88/53 Manfred KETS DE VRIES
88/54 Lars-Hendrik RULERand Mihkel M. TOMBAK
88/55 Peter BOSSAERTSand Pierre MILLION
88/56 Pierre MILLION
88/57 Vilfried VANHONACKERand Lydia PRICE
88/58 B. SINCLAIR-DESCAGNEand Mihkel M. TOMBAK
88/59 Martin KILDUFF
88/60 Michael BURDA
88/61 Lars-Hendrik RULER
88/62 Cynthia VAN HULLE,Theo VERNAELEN andPaul DE VOUTERS
'Whatever happened to the philosopher-king: theleader's addiction to pover, September 1988.
*Strategic choice of flexible productiontechnologies and welfare implications•,October 1988
•Method of moments teats of contingent claimsasset pricing models', October 1988.
"Size-sorted portfolios and the violation ofthe random valk hypothesis: Additionalempirical evidence and implication for testsof asset pricing models', June 1988.
'Data transferability: estimating the responseeffect of future events based on historicalanalogy", October 1988.
'Assessing economic inequality", November 1988.
"The interpersonal structure of decisionmaking: a social comparison approach toorganizational choice", November 1988.
"Is mismatch really the problem/ Some estimatesof the Chelvood Cate II model vith US data",September 1988.
'Modelling cost structure: the Bell Systemrevisited', November 1988.
' Regulation, taxes and the market for corporatecontrol in Belgium", September 1908.
"The impact of language theories on VSSdialog• . January 1989.
"DSS software selection, a multiple criteriadecision methodology". January 1989.
'Negotietiom support: the effects of computerinterventian and confliet level on bargainingoutcomes. January 1989.'Lasting improvement tomanuEsetorIngperformaace: Im search of • mew theory",January 1989.
'Shared history or shared culture? The effectsof time, culture, and perfernanee oninstitetionalization in simulatedorganisations*, January 1989.
°Coordiaating manufacturing and businessstrategies: February 1989.
'Structural adjustment fa Barone= retailtanking. Some view from industrialorganisation", January 1989.
'Trends in the development of technology andtheir effects on the production structure inthe European Community', January 1989.
"Brad proliferation and entry deterrence",February 1989.
'A market based approach to the valuation ofthe assets in place and the grovthopportunities of the firs', December 1988.
89/01 Joyce K. DIRER andTavfik JELASSI
89/02 Louts A. LE BLANCand Tavfik JELASSI
89/03 Beth 8. JONES andTavfik JELASSI
89/04 Kasra FERDOVS andArnoud DE METER
89/05 Martin KILDUFF andReinhard ANGEIJIAR
89/06 Mihkel N. TOMBAK andB. SINCLAIR-DESCAGNE
89/07 Damien J. NEVEN
89/08 Arnaud DE METER andRelimut SCRUM
89/09 Damien NEVEN.Carmen MATURES andMarcel CORSTJENS
89/10 Nathalie DIERKENS,Bruno GERARD andPierre BILLION
89/111 Manfred KETS DE VRIESend Alain NOEL
89/12 Untried VANUONACKFA
89/13 Manfred KETS DE TRIES
89/14 Reinhard ANGELMAR
89/15 Reinhard ANGELMA1
89/16 Vilified YAHOO/METER,Donald LEI:KANN andr SULTAN
89/17 Cilles APIADO,Claude FAUCNCUf andAndrd LAURENT
89/18 Scini IIALAK-RIS8MAN andMitchell KOLA
89/19 VilfrIed VANVONACKEA,Donald LERMANN and► SULTAN
69/20 Vilfried VAIDIONACKEAand Russell VIPER
89/21 Arnoud de METER andKau& FERMIS
89/22 Manfred KITS DC VRIESand Sydney PERZOV
69/21 Robert KORAJCZFK andClaude VIALLET
19/2a martin KILDUFF andMitchel AROLAFIA
09/25 Roger *ETAT:COURT andDavid GAUTSCHI
69/26 Charles SEAN,Edmond KALINVAUD,Peter BEINNOLZ,Francesco GIAMATTIand Charles VTPIOSZ
•Understanding the leader- tttttt a interface:applicAtios of the strategic relationshiplestroley netho • , rebruary 1989.
•Istlaettas dynamic response models otsen thedata are subject to differeet teApormlaggregation', January 1989.
•The Impostor syndrome: a disquietingpbeneratbon in organisational life". February1989.
'Product innovations • tool for coopetitiveadvantage', March 1989.
'Evaluating a firu's product innovationperformance*, March 1989.
'Coe:bluing related and sparse data in linearregression models*. February 1989.
• Changeeent organisationnel et rfallt4sculturelloss comtraates franco-meariceins..March 1989.
• Information asymmetry, market failure andjoint-ventures: theory and evidence',March 1989
•Combining related and sparse data in linearregressloo models••Revised March 1989
• A rational rondo. behavior model of choice'.Revised March 1989
• Imfluenee of nanufecturing improve...ear mores on performance'. April 1989
•Shot is the role of character inpsychoanalysts? April 1989
'Ilquity risk priests and the pricing of foreignexchange risk• April 1989
'The social destruction of reality:Organisational conflict es social duneApril 1969
•Too loll characteristics of retailmarkets and their economic consequences'March 1989
• Macroeconomic policies for 1992: thetransition and after', April 1989
89/27 David KRACKHARDT andMartin KILDUFF
89/28 Martin KILDUFF
89/29 Robert COCEL andJean-Claude LARRECHE
89/30 Ltrs-Hendrik ROLLERand Mihkel M. TOMBAK
89/31 Michael C. BURDA andStefan GERLACH
89/32 Peter HAUL andTavfik JELASSI
89/33 Bernard SINCLAIR-DESGAGNE
89/34 Sumantra GHOSHAL andNittin NORRIA
89/35 Jean DERMINE andPierre BILLION
89/36 Martin KILDUFF
89/37 Manfred KETS DE VRIES
89/38 Manfrd KETS DR VRIES
89/39 Robert KORAJCZTK andClaude VIALLET
89/40 Balaji CRAKRAVARTHT
89/41 B. SINCIAIR-DRSCAQIEand Nathalie DIERKENS
89/42 Robert ANSON andTavfik JELASSI
89/43 Michael RURDA
89/44 Balaji CBAKRAVARTHTand Peter LORANCE
'Friendship patterns and cultural attributions:the control of organizational diversity,April 1989
'The interpersonal structure of decisionmaking: • social comparison approach toorganizational choice• , Revised April 1989
•The battlefield for 1992: product strengthand geographic coverage • , May 1989
•Competition and Investment in FlexibleTechnologies', May 1989
' Durables and the US Trade Deficit', May 1989
'application end evaluation of a multi-criteriadecision support system for the dynamicselection of U.S. manufacturing locations•,May 1989
' Design flexibility in nonopsonisticindustries• , May 1989
'Requisite variety versus shared values:managing corporate-division relationships inthe M-Fore organisation• , May 1989
'Deposit rate ceilings and the market value ofhanks: The case of France 1971-1981 • , May 1989
'A dispositional approach to social netvorks:the case of organizational choice', Nay 1989
'The organisational fool: balancing • leader'shubris• , May 1989
"The C60 blues • , June 1989
' an empirical investigation of internationalasset pricing•, (Revised June 1989)
' Management systems for innovation andproductivity', June 1989
'The strategic supply of precisions', June 1989
"A development framework for computer-supportedconflict resolution', July 1989
'A note on firing costs and severance benefitsin equilibrium unemployment', June 1989
' Strategic adaptation in multi-business firms•,June 1989
89/45 Rob VEITZ and
"Managing expert systems: a framevork and easeArnoud DE METER study*, June 1989
89/46
89/47
Marcel CORSTJENS, 'Entry Encouragement', July 1989
Carmen MAIMS and
89/64(TM)
Enver YUCESAN andLee SCHRUBEN
"Complexity of simulation models: A graphtheoretic approach", November 1989
Damien NEVER89/65 Soumitra DUTTA and "MARS: A mergers and acquisitions reasoning
Manfred KETS DE VRIES 'The global dimension in leadership and
and Christine MEAD organization: issues and controversies',(TM,AC, PIN)
Piero BONISSONE system", November 1989
April 1989
89/48 Damien NEVEM andLars-Bendrik ROLLER
89/49 Jean DERMINE
89/50 Jean DERMINE
89/51 SpyrosMAKRIDAKIS
89/52 Arnoud DE METER
89/53 Spyros MAXRIDAKIS
89/54 S. BALAKRISHNAN
and Mitchell KOZA
89/55 H. SCHUTTE
89/56 Vilfried VANHONACKPRand Lydia PRICE
89/57 Taekvon KIM,
Lars-Bendrik ROLLERand Mihkel TOMBAK
89/58 Lars-Hendrik ROLLER(EP,TM) and Mihkel TOMBAK
89/59 Manfred KETS DE VRIES,
(08) Daphne ZEVADI,
Alain NOEL and
Mihkel TOMBAK
89/60 Enver YUCESAN and(TM) Lee SCHRUBEN
89/61 Susan SCBNEIDER and(All) Arnoud DE METER
89/62 Arnoud DE MEYER
(TM)
89/63 Enver TUCESAN and
(TM) Lee SCRRUBEN
•European integration and trade flows',
August 1989
'Home country control and mutual recognition',
July 1989
"The specialization of financial institutions,
the EEC model", August 1989
'Sliding simulation: a new approach to time
series forecasting', July 1989
'Shortening development cycle times: a
manufacturer's perspective', August 1989
'Sly combining works?', July 1989
'Organisation costs and ■ theory of joint
ventures', September 1989
'Euro-Japanese cooperation in information
technology', September 1989
'On the practical usefulness of ■eta-analysis
results', September 1989
'Market growth and the diffusion ofmultiproduct technologies', September 1989
'Strategic aspects of flexible productiontechnologies', October 1989
'Locus of control and entrepreneurship: a
three-country comparative study', October 1989
'Simulation graphs for design and analysis of
discrete event simulation models', October 1989
'Interpreting and responding to strategicissues: The impact of national culture',
October 1989
'Technology strategy and international R 6 D
operations', October 1989
'Equivalence of simulations: A graph theoretic
approach • , November 1989
"On the regulation of procurement bids",November 1989
"Market microstructure effects of government
intervention in the foreign exchange market",December 1989
89/66 B. SINCLAIR-DESGAGNE(TM,BP)
89/67 Peter BOSSAERTS and(PIN) Pierre BILLION
1990
90/01/TM B. SINCLAIR-DESGAGNE "Unavoidable Mechanisms", January 1990EP/AC
90/02/EP Michael BURDA "Monopolistic Competition, Costs ofAdjustment, and the Behaviour of EuropeanManufacturing Employment", January 1990
"Management of Communication in InternationalResearch and Development", January 1990
"The Transformation of the European FinancialServices Industry: From Fragmentation toIntegration", January 1990
"European Equity Markets: Toward 1992 andBeyond", January 1990
"Integration of European Equity Markets:Implications of Structural Change for KeyMarket Participants to and Beyond 1992",January 1990
"Stock Market Anomalies and the Pricing ofEquity on the Tokyo Stock Exchange", January1990
"Modelling vith MCDSS: What about Ethics?",January 1990
"Capital Controls and International TradeFinance", January 1990
"The Impact of Language Theories on MSDialog", January 1990
"An Overview of Frequency Domain Methodologyfor Simulation Sensitivity Analysis",January 1990
90/03/TM Arnoud DE MEYER
90/04/ Gabriel HAWAWINI andPIN/EP Eric RAJENDRA
90/05/ Gabriel HAWAWINI andFIN/EP Bertrand JACOUILLAT
90/06/ Gabriel HAWAWINI andFIN/EP Eric RAJENDRA
90/07 Gabriel HAWAWINIFIN/EP
90/08/ Tavfik JELASSI andTM/EP B. SINCLAIR-DESGAGNE
90/09/ Alberto GIOVANNINIEP/FIN and Jae VON PARK
90/10/TM Joyce BRIER andTavfik JELASSI
90/11/TM Enver YUCESAN
90/12/EP Michael BURDA
"Structural Change, Unemployment Benefits andHigh Unemployment: A U.S.-EuropeanComparison", January 1990