Creating Social Agent Systems in Simulated Physical Environments

Principal Investigator: Dr. Ben AguirreResearch Assistant: Eric Best

(ericbest@udel.edu)

Lead, Follow, or Get in the Way

1

Why Social Agent Systems?

• Most simulations create a number of avatars that are largely homogeneous and do not have linked goals.

• Based on our research, we believe this is an inaccurate way to create models of human beings.

2

Disaster Research Center Work

• Based on our work at the Disaster Research Center (DRC), we have found that people evacuating a hazard do not rely on individual goals and behaviors.

• Group behaviors and crowd behaviors impact goals and reasoning.

3

Why Group Behavior?

• Historically, the DRC has found that evacuees use social behavior when assessing and reacting to hazards.

• Social bonds are not ignored when dangers are present.

• Individual, social, and crowd behaviors interact together to influence decisions and goals.

4

Why Group Behavior?

• Recently, Dr. Ben Aguirre conducted a study that gathered detailed information about the victims and evacuees of The Station Nightclub fire.

• We are aware of detailed exit goals, social behaviors, and networks among victims and evacuees.

5

Why Group Behavior in ABM?

• Using this detailed dataset, we chose to create a set of models incorporating social behavior to replicate the immediate reaction to the disaster at The Station.

• Eventually, we hope to encourage other model builders to consider group behavior and non-homogenous agents.

6

Why Group Behavior in ABM?

Humans do not interact in a vacuum, but most simulations of humans account only for individual behavior.

Agent-Based modeling with learning agents is an obvious avenue for creating simulations with evolving group behavior models.

7

Our Agent-Based Models

• Model builders are largely ignoring group and crowd behaviors, so we set out to build a model incorporating social behavior models.

• Today, we will discuss some of the conceptual and practical issues that must be addressed when creating models with group behavior attributes.

8

In Practice

One of several first efforts in physical social agent simulation

Several benchmarks

9

In Practice

• Goals for model creation:

– Benchmark current models and actual results.

– Analyze data and create an individual behavior model.

– Create models with group behavior attributes.

– Apply framework to other situations.

10

Benchmarking

• For The Station, we have a detailed dataset with results that we can compare to output from models of the disaster.

• There are several models that have been created about The Station, many with a different focus, but none with social behavior.

• We have exit data and exit time data. Most results compare flow rates in aggregate.

11

Analysis

• The DRC conducted a study gathering highly detailed data of victims and evacuees, including social bonds to other participants.

• Using this data, we were able to approximate the likelihood of behavior changes based on different social bonds.

12

Individual Behavior Model

• Like many others, we created a model with individual behavior.

• This became our personal benchmark. If we can control all of the parameters, we can see the exact impacts of adding group behavior traits.

• Our individual behavior model has similar flow rates to data results and prior models.

13

Background

• We began with an issue of geography. We had to create an accurate environment of The Station. For now in two dimensions.

• We then added hazards, reasons to leave the building.

• We then created our avatars, or agents.

14

Agents

• Our avatars have different behaviors based off of observed traits, such as age, gender, familiarity with the environment, and other factors. They are not homogenous.

• Every avatar has a large set of constraints, such as the inability to move through walls.

• Avatars also have to navigate in the changing environment due to other avatars moving.

15

Adding Behavior Layers

• Once our individual behavior model was running, we added other layers:

• Individual layer.

• Social group layer.

• Crowd layer.

16

Adding Social and Group Behaviors

• After creating an individual behavior model, we began to add group elements.

• Almost immediately, our efficiency plummeted, avatars stopped exiting, and our results made little sense when examined by a human.

17

What Happened?

• It is easy to forget that computers are not nearly as smart as you are.

• Every time you add a variable to an agent's decision tree, you have to account for interactions.

• Conflicting goals will render a machine's reasoning system useless unless they are programmed to perform an action or prefer a result.

18

What Happened?

• The moment the actions of other avatars matter to an agent, a high number of interaction variables emerge.

• A simple first-order interaction of 500 agents creates 125,250 unique relationships.

• Adding additional complications can increase interactions exponentially.

19

Where to begin?

Before you can create rules or preferences, you have to identify the issues.

The simplest interactions can be hard to quantify.

20

Building a Model

• There is a multi-step process of conceptualization, implementation, and robustness testing.

• This process is inherently interdisciplinary.

21

Complications

These physical issues quickly become difficult computational and judgement problems.

Simple group movements require constant recalculation

22

Complications

• Individual, group, and crowd behaviors create large numbers of scheduling issues and interactions that must be understood.

23

Complications

• Reasoning that humans take for granted is often difficult to implement in computerized systems.

• It is not feasible to truly simulate humans, so it becomes important to focus on a few programmable traits.

• How are social traits programmable?

24

Complications

• Here is an example meant to relay the concept of object permanence.

• How do you teach this to a computer, or even communicate it to a programmer?

25

An Interdisciplinary Approach

• Identifying and coding behaviors merges ideas and techniques from sociology, psychology, computer science, economics, and statistics among others.

• The DRC is uniquely positioned to accept these challenges thanks to strong interdisciplinary programs and expertise sharing.

26

Where Does ABM Fit In?

Agents that can interact together are a natural fit for physical environment simulations. In theory, they just require one more layer of rules to create a physical reality.

In practice, group behavior interactions and physical limitations are going to keep a young field busy for quite awhile. Humans are more difficult to simulate than one might imagine.

27

How Do We Make Working Models?

• Hurdles:

– High-Performance Computing: Relationships tax computing resources. Thankfully, we now have parallel systems and GPU computing.

– Peer Review: Designers are reluctant to share code, and even if shared, it is often difficult to understand. Debugging is very time intensive.

– What attributes to code: Who decides what aspects of group behavior are important?

28

Where Are We Today?

• Currently, I spend most of my time tackling the third issue; which traits are most important to include.

• To begin, I created a version of our model with a simple leader/follower group model.

• These groups are small social dictatorships, where followers pattern a leader exhibiting an individual behavior model.

29

Where Will We Go?

• Our eventual goal is to implement distributed social behavior models, approximating complex social relationships.

• These models will have decentralized and robust leadership, much like networked computer systems.

• Beyond groups, we will incorporate crowd models.

30