Predator-prey simulation using MARL

   page       attach   
Luca Fabri
abstract

Reinforcement Learning (RL) is a sub-area of Machine Learning which focuses on building strategies to make decisions that maximize the future expected reward. The RL finds applications in problems concerning self-driving cars, industrial automation, and finance but also in systems that involve the interaction of multiple agents in a shared environment, where it’s more specifically called Multi-Agent RL (MARL). In computational biology, it’s useful to study the population of one or more intelligent species that interact with each other to build population models: in this project, a predator-prey ecosystem is implemented, using a MARL approach in a mixed environment, i.e. cooperative and competitive, where agents of the same species cooperatively make decisions to maximize their total expected reward. To achieve this goal, the MADDPG algorithm is exploited. The distribution of the system is realized by parallelizing the environments and by introducing a distributed training technique, inspired by the Mava framework.

outcomes