EvoPlaSNN: Evolving Reward-modulated ANN-based Plasticity Rule for Spiking Neural Networks
Neuroevolution can indirectly train a neural network by optimising the underlying plasticity rules that adapt the network weights. In Spiking Neural Networks (SNNs), evolvable meta-learning studies have been applied to unsupervised learning settings. In our experiment, we let evolution search for rules that can train an SNN to solve a Reinforcement Learning maze. These rules are encoded as an Artificial Neural Network (ANN), taking as inputs synaptic variables such as weight, reward and eligibility traces. We found that using synaptic weight as inputs has no effect on rule performance, while utilising either pre-before-post or post-before-pre eligibility trace is better than a trace that combines both. Although there is still too much variation in fitness both within individuals and across generations to allow for conclusive results, there is potential for evolving plasticity rules to solve RL tasks in the future.