Catgame: A Tool For Problem Solving In Complex Dynamic Systems Using Game Theoretic Knowledge Distribution In Cultural Algorithms, And Its Application (catneuro) To The Deep Learning Of Game Controller
Open Access Dissertation
Date of Award
Robert G. Reynolds
Cultural Algorithms (CA) are knowledge-intensive, population-based stochastic optimization methods that are modeled after human cultures and are suited to solving problems in complex environments. The CA Belief Space stores knowledge harvested from prior generations and re-distributes it to future generations via a knowledge distribution (KD) mechanism. Each of the population individuals is then guided through the search space via the associated knowledge. Previously, CA implementations have used only competitive KD mechanisms that have performed well for problems embedded in static environments. Relatively recently, CA research has evolved to encompass dynamic problem environments. Given increasing environmental complexity, a natural question arises about whether KD mechanisms that also incorporate cooperation can perform better in such environments than purely competitive ones? Borrowing from game theory, game-based KD mechanisms are implemented and tested against the default competitive mechanism – Weighted Majority (WTD).
Two different concepts of complexity are addressed – numerical optimization under dynamic environments and hierarchal, multi-objective optimization for evolving deep learning models. The former is addressed with the CATGame software system and the later with CATNeuro.
CATGame implements three types of games that span both cooperation and competition for knowledge distribution, namely: Iterated Prisoner's Dilemma (IPD), Stag-Hunt and Stackelberg. The performance of the three game mechanisms is compared with the aid of a dynamic problem generator called Cones World. Weighted Majority, aka “wisdom of the crowd”, the default CA competitive KD mechanism is used as the benchmark. It is shown that games that support both cooperation and competition do indeed perform better but not in all cases. The results shed light on what kinds of games are suited to problem solving in complex, dynamic environments. Specifically, games that balance exploration and exploitation using the local signal of ‘social’ rank – Stag-Hunt and IPD – perform better. Stag-Hunt which is also the most cooperative of the games tested, performed the best overall. Dynamic analysis of the ‘social’ aspects of the CA test runs shows that Stag-Hunt allocates compute resources more consistently than the others in response to environmental complexity changes. Stackelberg where the allocation decisions are centralized, like in a centrally planned economic system, is found to be the least adaptive.
CATNeuro is for solving neural architecture search (NAS) problems. Contemporary ‘deep learning’ neural network models are proven effective. However, the network topologies may be complex and not immediately obvious for the problem at hand. This has given rise to the secondary field of neural architecture search. It is still nascent with many frameworks and approaches now becoming available. This paper describes a NAS method based on graph evolution pioneered by NEAT (Neuroevolution of Augmenting Topologies) but driven by the evolutionary mechanisms under Cultural Algorithms. Here CATNeuro is applied to find optimal network topologies to play a 2D fighting game called FightingICE (derived from “The Rumble Fish” video game). A policy-based, reinforcement learning method is used to create the training data for network optimization. CATNeuro is still evolving. To inform the development of CATNeuro, in this primary foray into NAS, we contrast the performance of CATNeuro with two different knowledge distribution mechanisms – the stalwart Weighted Majority and a new one based on the Stag-Hunt game from evolutionary game theory that performed the best in CATGame. The research shows that Stag-Hunt has a distinct edge over WTD in terms of game performance, model accuracy, and model size. It is therefore deemed to be the preferred mechanism for complex, hierarchical optimization tasks such as NAS and is planned to be used as the default KD mechanism in CATNeuro going forward.
Waris, Faisal, "Catgame: A Tool For Problem Solving In Complex Dynamic Systems Using Game Theoretic Knowledge Distribution In Cultural Algorithms, And Its Application (catneuro) To The Deep Learning Of Game Controller" (2020). Wayne State University Dissertations. 3412.