Using Genetic Algorithim and Neural Network for learning and adaptive behavior in a goal finding application

Students: James Stigall

Past Students: Tatwamasi Tripathy, Taiysha Hobbs,Kola Ogunlana

Status: Current

The goal of this project is combining Genetic Algorithm (GA) with Neural Networks (NN) to explore how intelligent agents can look for exits during an evacuation. The agents have the capability to adapt their behavior in the environment and formulate their response by learning from the environment. Our approach focuses on modeling individual behavior as well as group behavior. Individuals constantly adjust their behavior according to the dynamic factors in the environment. We have developed crowd-modeling and emergency behavior modeling capability in a goal finding application using C#. This project examines an intelligent agent-based evacuation that can help plan emergency evacuations, run numerous event-driven evacuation scenarios, support research in the areas of human behavior, and model the movement of responders and security personnel. The adaptive behavior focuses on the individual agents changing their behavior in the environment. The collective behavior of the agent focuses on the crowd-modeling and emergency behavior in the goal finding application. We have developed a new intelligent agent based characteristics such as autonomy, social ability, cooperativeness, and learning ability which define their final behavior when trying to reach a goal. Our contributions lies in our approach of combining Genetic Algorithm (GA) and Neural Network (NN) to model learning and adaptive behavior of agents in a goal finding application. The proposed application will aid in running multiple evacuation drills for what if scenarios by incorporating agent characteristics.


A hierarchy of behaviors and neural network design

An Agent moving towards the goal


In our proposed goal finding application, the agents are classified in two types as shown in above figure: 1) Steering Agent 2) Intelligent Agent. We have built a goal finding application using C# for crowd movement where the agents are shown in different colors. When the simulation starts the agents have an assigned initial goal. Some of these agents have a secondary goal that leads them to group together as they move towards the goal.


 Path following behavior

Intelligent agents going towards the goal

The key aspect of agent based modeling lies in understanding the mechanism by which autonomous agents interact among themselves. Our hypothesis is that people with similar background, race, ethnicity, sex etc are more likely to collaborate together and exhibit close affinity in emergency than people with varied background. Agents come from different background with the capability to adapt their behavior under different environment and formulate their response by learning from the environment. The agents have characteristics such as:

  1. Attributes: An agent is a discrete individual and has mass, position, velocity, force, and speed.
  2. Emotions: An agent has emotions such as level of panic and stress attributes.
  3. Memory: The agents are goal oriented. The agents have a list of goals that constantly keeps increasing or decreasing depending upon its interaction with the environment.
  4. Rules of behavior: An agent has the ability to learn and adapt its behaviors based on experience in the environment using genetic algorithm and neural network.
  5. Decision making capability: An agent is autonomous and can function independently in its environment while interacting with other agents (dynamic obstacles) and environment (static obstacles).

Our proposed multi agent system models human behavior. The agent characteristics include attributes, emotions, memory, rules of behavior and learning, and decision making capabilities. Our proposed multi agent system also uses MLP (refer fig. below) and consists of variables which provide a way to keep existing state as it moves thereby relying on the stimulus it receives from the world.

Multi Perceptron Network Agents evacuating in a multi-agent system

Our proposed C# application can be used to model situations that are difficult to test in real-life due to safety considerations. It is able to include agent characteristics and behaviors. The findings of this modeling are very encouraging as the agents are able to assume various roles to combine GA and NN on the way to reaching their goals.

We are currently integrating fuzzy logic to model stress, panic, and the uncertainty of emotions. The findings of this modeling are very encouraging as the agents are able to assume various roles to utilize fuzzy logic on the way to reaching their goals. The fuzzy rules link the parts together while feeding into behavioral rules.