Transfer of driving behaviors across different racing games

Transfer learning might be a promising approach to boost the learning of non-player characters' behaviors by exploiting some existing knowledge available from a different game. In this paper, we investigate how to transfer driving behaviors from The Open Racing Car Simulator (TORCS) to VDrift, which are two well known open-source racing games featuring rather different physics engines and game dynamics. We focus on a neuroevolution learning framework based on NEAT and compare three different methods of transfer learning: (i) transfer of the learned behaviors; (ii) transfer of the learning process; (iii) transfer of both the behaviors and the process. Our experimental analysis suggests that all the proposed methods of transfer learning might be effectively applied to boost the learning of driving behaviors in VDrift by exploiting the knowledge previously learned in TORCS. In particular, transferring both learned behaviors and learning process appears to be the best trade-off between the final performance and the computational cost.