Reinforcement learning for ecosystems- Using explainable dynamic neural networks to train reinforcement learning agents in a simulated virtual animal ecosystem

Abstract

This report covers the development of an inherently explainable and dynamic neural network for reinforcement learning in virtual animal ecosystems, based on the lifelong learning from zero (LL0) neural network used in supervised learning. The developed network (RLL0) is a fuzzy neural network with specialised growing and pruning rules for an ever changing environment. Results from benchmarking against a reference network show that RLL0 has a comparable performance while using far fewer trainable parameters. This, combined with its adapted architecture for visualising a learnt behaviour, shows promising future extensions to and use of the explored algorithms for animal behavioural based biological simulations.