Evolving Transferable Artificial Neural Networks for Gameplay Tasks via NEAT with Phased Searching

NeuroEvolution of Augmenting Topologies (NEAT) has been successfully applied to intelligent gameplay. To further improve its effectiveness, a key technique is to reuse the knowledge learned from source gameplay tasks to boost performance on target gameplay tasks. We consider this as a Transfer Learning (TL) problem. However, Artificial Neural Networks (ANNs) evolved by NEAT are usually unnecessarily complicated, which may affect their transferability. To address this issue, we will investigate in this paper the capability of Phased Searching (PS) methods for controlling ANNs’ complexity while maintaining their effectiveness. By doing so, we can obtain more transferable ANNs. Furthermore, we will propose a new Power-Law Ranking Probability based PS (PLPS) method to more effectively control the randomness during the simplification phase. Several recent PS methods as well as our PLPS have been evaluated on four carefully-designed TL experiments. Results show clearly that NEAT can evolve more transferable and structurally simple ANNs with the help of PS methods, in particular PLPS.