Physically Grounded CBR By Constantin Savtchenko.

24
Physically Grounded CBR By Constantin Savtchenko

Transcript of Physically Grounded CBR By Constantin Savtchenko.

Physically Grounded CBR

ByConstantin Savtchenko

CBR As We Know It

• So far, we’ve looked at– Expert systems• Make good decisions and solutions in one specific

domain

– Recommender Systems• Aggregate data and suggest similar cases

– Game AI• Make in-game decisions in order to beat an opponent

• Overarching property: abstract domains!

Physically Grounded CBR

• The physical world is the domain• Requires the ability to measure the current

state of environment– Distances– Colors– Temperature– Bumpers

• The field of robotics operates in this domain, and some applications use CBR

Intro to Mobile Robotics

• We’ve already discussed some basic robotics with Prof. Munoz (ie. factory automation)

• Now we look at autonomous mobile robotics– Autonomous– Mobile

• Three key questions– Where am I?– Where am I going?– How do I get there?

• A CBR has been applied to answer each question!

Basics of Mobile Robotics

• Typically need the following 4 states to answer the three questions

Perception(What do I see?)

Localization(Where am I?)

Cognition(Whats the next step?)

Motor Control(How do I move?)

Real World

Difficulties Of Mobile Robotics

• Uncontrollable nature of the world• Imperfect measurements– Incorrect perception– Incorrect movements

• Infinitely large domain– How do we account for everything?

• Split second decisions– There is no time to debate

Solving The Difficulties Of Mobile Robotics

• Uncontrollable nature – make predictions• Imperfect measurements– Use a priori knowledge– Take multiple measurements

• Infinitely large domain – make generalizations• Real time decisions – take shortcuts/heuristics• With these basics in mind, we can begin to

represent our problems

Designing a CBR System for Mobile Robotics

• Case Representation – The hardest and most important part– How should we define cases? Domain is infinite and time based

• How do we determine case similarity in continuous environments?– Metrics we know– Logic to reduce case retrieval overhead

• What is the scope of my solution?– Low level action?– High level mentality?

• Success of Solution– How do we judge the success of the solution?

• Autonomous Robotics – no operator to guide– Low tolerance for bad decisions– Split second decisions

• What part of mobile robotics are we going to deal with?

CBR - Where Am I?

• A CBR system was created to use landmarks to describe unknown environments

• By using a priori knowledge, robot can make informed decisions– Moving into a corner– Predicting walls

• Similarity function used landmarks between current situation and cases

• Ros, R. Lopez de Mantaras, R. Sierra, C. Arcos, J, L. A CBR System for Autonomous Robot Navigation. Spanish council for Scientific Research.

CBR – Where Am I Going

• AIBO Robotic dogs are used to compete in Robocup– Soccer for robotic dogs

• Case-base reasoning usedto determine game playswhich are high level actiondecisions– Dribble and shoot

AIBO Robocup – The Nitty Gritty• Case Description

– Robot Position– Ball Position– Defending Goal– Teammate Position– Opponent Position– Current Time– Score difference

• Solution Description– Sequence Of Actions

• Similarity Metrics– Gaussian distribution

of any real values (a)– Simple symbolic

equivalence for goal (b)– A piecewise function

to describe the strategy (c)

Controllable

Controllable

Not Controllable

Not Controllable

Not Controllable

Not ControllableNot Controllable

AIBO Robocup – Making A Decision• Case Retrieval

– Find similar cases based on non controllable features.• Important design decision, single out cases that are ABSOLUTELY similar. Then work from

there

– So what’s the points of controllable features?• Minimize number of cases• Create a cost function between cases – how much work would it take to modify my current

situation?

• Case Adaptation – Cost– Measures the difference between controllable euclidean distances (team robot’s

positions, in the current situation and the case)–

• Maximize similarity but minimize cost– – Notice that we move away from picking the best decision, and towards the more

efficient – This helps nullify any consequences of a bad solution

CBR – Where am I going?Advanced Strategies

• That didn’t seem too high level, it was just smart planning… Right! But it was a great start.

• Case representation was modified to give cases “scope” or inherent similarity based on a “general area”

• Actions were changed to involve multiple robots performing sequences of low level actions– Robot 1: dribble, pass to Robot 2

Robot 2: move forward, wait, shoot• Working with the basics from before, the research group

changed their CBR to implement team-strategies!– A huge jump in logic is made in a very simple manner since the base

CBR system already existed!

Changes in the Basic CBR System

• A Knowledge Base is added, this is heuristically important information, that helps find cases– Instead of comparing

actual positions, focuson general positions.We are more interested in qualitative positions. Thus (a) and (b) are actually the same situation

• When making a decision, maximize similarity, minimize cost AND maximize number of players in game play

Gaining A Lot By Doing a Little

• The perfect situation!• Using previous basic technology, simple

modifications were made to create very high level strategies– Solutions are sequences of (robot, action) pairs.– Actions are still very basic

• Results?

CBR - How do I get there?

• We just saw a great example of how to discretize our time based on input of a continuous domain.– What happens when we summarize the last 5 seconds?

• We need to make instant decisions, as we move, can’t summarize!

• Continuous Case Base Reasoning– Continuous representation – cannot discretize– Continuous performance – Decisions must be made at any moment– Continuous Adaptation and Learning – An autonomous agent must

adapt as problems get more difficult• To achieve these requirements, we can’t use discrete

symbolic reasoning systems.

CBR – Mobile Robotic Navigation• Basics of Navigation:

– Successfully navigating from the current location to a goal location in an obstacle ridden terrain

– Do we plan ahead of time?• How much should we plan for?• What happens if we do not foresee everything?

– Do we just wing it and react as things come up?• How do we know we’re not going in circles?• How do we know we didn’t just make an even worse decision?

• Can we put the two systems together?– Yes! Lets put low level behaviors together into groups of behavior

assemblages:• Aggressive behaviors together, move quickly with a high goal attraction• Cautious behaviors together, move slowly cautiously move around obstacles

Behavioral Navigation• Reactive:

– Behaviors is determined by parameters• Limit on how close we can get to an obstacle• Limit on speed

• Planning:– Modify parameters slowly as we navigate,

but in a suddenly new environment, make a huge change!• Navigate through doorway, then you’re outside

• Results:– A general plan on how to handle different environments which adjusts as

new input comes in.– Requires us to continuously be making changes. Enter

continuous case representations

Continuous Case Base Reasoning

• How do you describe a continuous case!?!– Make a vector of the

value of each continuous value in terms of time.

• Similarity– Similarity measured by

mean squareddifference between vector representationof the environment.

CBR And Perception

• VADER Lab: The Smart Wheelchair Project– Create a wheelchair that can navigate

environments autonomously given a destination command

• How to avoid people 101– See some object– Determine if it’s a person– Determine motion path if it is a person

Determining People Using Case-Based Reasoning

• People are not the same shape, nor do they look the same from all sides– Makes a non unimodal distribution– Difficult to define everything possible and react to each

situation• The measurements of people is in continuous space

– Difficult arises when setting thresholds.– High overlap between people and non-people

• Solution: Nearest Neighbors– Can deal with non unimodal data– Decisions can be done on similarity rather than thresholds

Advantages of Case-Base Reasoning

• Easy to modify case representation and instantly see results

• Many applications only use retrieve and reuse, however, easy and beneficial to implement retain– The more retained, the more accurate the reasoner

• Very simple implementation, less time spent coding, more on research of features.

• Quick and simple high level decisions that use groups or assemblages of low level actions/behaviors

• Easy to understand and maintain behavior of robots due to representation of cases

Disadvantages of CBR in Mobile Robotics

• Very high-level solutions. Robotics typically requires a very good set of low level, PRECISE, solutions

• Can be rigid at first until the Case Base and actions have been expanded upon.

Bibliography• Ram A. & Santamaria J.C. Continuous Case-Based Reasoning. In Proc. Of

the 1993 AAAI Workshop on Case-Based Reasonings, pp. 86-93, Washington, DC, 1993

• Ros R., Lopez de Mantaras R., sierra C., Arcos J.L. A CBR system for autonomous robot navigation. From Artificial Intelligence Research Institute.

• Ros R., Veloso M., Lopez de Mantaras R., Sierra C., Arcos J.L. Retrieving and Reusing Game Plays for Robot Soccer. In Roth-Berghofer, T.R., Goker, M.H., Guvenir, H.A. (eds.) ECCBR 2006. LNCS (LNAI), vol. 4106, pp. 47–61. Springer, Heidelberg, 2006

• Ros R., Lopez de Mantaras R., Arcos J.L., Veloso M. Team Playing Behavior in Robot Soccer: A Case-Based Reasoning Approach. In R.O. Weber and M.M. Richter (Eds.): ICCBR 2007, LNAI 4626, pp. 46–60, 2007