Neuro-evolution using game-driven cultural algorithms
Faisal Waris, Robert G. Reynolds · 2020
Contemporary 'deep learning' (DL) models have proven to be effective in a wide variety of applications. However, the right network topology for the problem at hand may be complex and not immediately obvious. This has given rise to the secondary field of neural architecture search (NAS). This paper describes a NAS method based on graph evolution pioneered by Neuro-evolution of Augmenting Topologies (NEAT), but driven by the evolutionary mechanisms underlying Cultural Algorithms (CA). CA is a population-based, stochastic optimization system inspired by problem solving in human cultures, suited to solving problems such as NAS. We present CATNeuro a system for evolving DL models guided by CA metaheuristics called Knowledge Sources (KS). The KS store knowledge harvested from prior generations and use it to guide subsequent generations in the search space. A knowledge distribution mechanism, which assigns a KS to each individual in the population, is an instrumental part of this process. CATNeuro, is applied to find optimal network topologies to play a 2D fighting game called FightingICE (based on "The Rumble Fish" game). A policy-based, reinforcement learning method is used to create the training data for network optimization. CATNeuro is still evolving. In this primary foray into NAS, we contrast the performance of CATNeuro with two different knowledge distribution mechanisms - the stalwart Weighted Majority (WTD) - which represents "wisdom of the crowds" - and a new one based on the Stag-Hunt game from evolutionary game theory. We show that Stag-Hunt has a statistical edge over WTD in many areas and thus is a better candidate for future development with CATNeuro.