Simulation of Homogenous Fish Schools in the Presence of Food and Predators using Reinforcement Learning

Abstract

We utilized Deep Reinforcement Learning to incor-porate schooling, foraging, and predator avoidance behaviors into a single fish behavior model. We used Proximal Policy Optimization (PPO) with Intrinsic Curiosity Reward (ICR) to make fish agents learn in our Unity Environment. We created an interactive control system on Unity that allows users to visualize and manipulate the simulation using only a mouse and keyboard. We compared our model with three variations: one without schooling reward, one without foraging reward, and one without predator avoidance reward. Our original model (schooling, foraging, and predator avoidance) clearly illustrated the unification of all three behaviors.