Simulating sentient agents

Kevin Galloway
CS 480

One important consideration when it comes to graphical simulations is the idea of intelligent thought. Particle engines, for example, have been used to simulate crowds that flee from giant monsters as well as to simulate traffic. One thing these two things have in common is human thought, people don't just flee in straight lines, or bounce off of each like pinballs.

A Framework for Sentient Agents

One way to solve this problem is to create a framework for simulating sentient behavior. Each timestep one would pass parameters of a thing (person, robot, Godzilla) into a rules engine that would decide on how that thing would move, that is, based on a creature's knowledge, it will react. If Godzilla sees the puny tank in front of it, he will crush it, but if he sees Mothra and Super Mechagodzilla, he might flee. This framework will be composed like so:
Population Generator -> Rule engine -> Visualizer

Population Generator

The population generator is one of the most important parts of this framework. This will create your objects, and give each parameters that will make them all similar, but still different. For example, your population generator might create people that are trapped in a building. Your class might look as follows:

Class Person{
int age;
int speed;
int aggression;
vec3 position;
float vision_range;
};

The population generator will fill these with random, but sensible values, so there will be no negative ages, nor ages above, say 100. These values will be placed into the rule engine.

Rule Engine

The Rule Engine has several different components to it, that simulate a variety of actions that a sentient agent can do. It is composed of a sensor , decision rules, and actuators. The sensor is how each sentient agent "sees" its world. One can have an agent have a ray that is cast ahead of them, and returns true if it hits something. Alternatively, one can just simply check to see if anything is near or collides with it. The sensor is incredibly important, as it makes each agent have an imperfect view of the world, so it does not know everying about the environment.

Decision rules are what happens if an agent sees something. This operates on a variety of factors: What does the agent see? Does it have any internal attributes (importance, aggression, tension) that it responds to? So if you have a person encounter Godzilla and he's very aggressive and angry, he might choose to attack, whereas someone not aggressive would flee.

Another portion of the Rule engine is the Global Database. This holds the information about every individual, and the environment. This is so a general feel of the crowd (the total crowd's tension or reaction for example ) can be used to modify an individual's behavior. One person fleeing from Godzilla our aggressive hero could ignore, 30 or 40 terrified people, and our hero might reconsider his attack.

Combining all of these is what creates the rules engine. Each agent steps through an iterative process: ask for a decision, get immediate environment (that is tension level/crowd density/immediate dangers), check for rules. These rules are all based on what independent attributes each agent has.

The Visualizer

To be honest, there isn't a lot to talk about the visualizer. This could be anything from points on a polygon, to an incredibly well rendered 3d game. Really, the sky's the limit here.

Pitfalls:

One thing that must be considered when doing a simulation of sentient agents is that all actions for every agent must be done "simultaneously". This does not mean that to do this, one must have parallel algorithms, but that all movements have to happen at the same time, followed by the rule engine. If one tries to couple movement and rules, then agents could be acting on incorrect data.

Designating the environment is incredibly important for this sort of simulation as well. If agents are operating on a faulty "understanding" of the world, then weird things can happen, such as walking through walls, if for example, an agent's speed is larger than the width of a wall.

Links:

http://eil.stanford.edu/egress/publications/ICCCB_paper.pdf
http://pdcc.ntu.edu.sg/dsgrid06/doc/0064_cao-trafficsimulation.pdf
Repast Symphony is a Java agent simulation framework and handy GUI.