Swarm learning in restricted environments: an examination of semi-stochastic action selection

This paper explores a machine learning process for robotic swarms tasked with a non-trivial problem in restricted environments. The effect of using a semi-stochastic action selector in a learning classifier based behaviour system is examined via adjusting the stochasticity setting. In this study we utilise Greedy Randomised Adaptive Search Procedures, finding some improvement in the ability of the swarm in non-deterministic, partially observable environments, compared to Greedy selection. We also find the swarm performs significantly worse when machine learning is removed. This study also explores an evolutionary process used to optimise the behaviours available to each agent. This evolutionary process is examined in regard to the effect it has on the learning settings. It is found that the evolution reduces the impact of fine-tuning the learning variables. However, fully stochastic selection prevents learning, which impairs the evolution.