Hybrid competitive-cooperative coevolution of decentralized controller in EvoTanks

• Main Author: Tan, Lawrence C.
• Format: text
• Language: English
• Published: Animo Repository 2008
• Subjects: Games > Data processing; Neural networks (Computer science); Coevolution; Machine Learning; Computational Intelligence; Games; Neural Networks; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_masteral/3476; https://animorepository.dlsu.edu.ph/context/etd_masteral/article/10314/viewcontent/CDTG004219_P.pdf
• Institution: De La Salle University

Competitive coevolution has often been the preferred choice for coevolving strategies in one-on-one games because it does not require a hand coded or pre-programmed solution to guide the search. While cooperative coevolution possesses the advantage of problem decomposition, it suffers from the requirement of an external or pre-programmed agent to guide the search towards its goal. This thesis extends previous works on cooperative coevolution by integrating it with competitive coevolution so that reliance on hand coded agents is eliminated. As a proof of concept, it is applied in EvoTanks (Thompson, 2006), a tank-based combat game. The hybrid implementation decomposes a three input three output network into three separate networks each with three inputs and one output. These decentralized neural networks are coevolved through a host-parasite cooperative coevolution setup and compared to a standard host-parasite competitive coevolution. Meta-test comparison revealed that the hybrid approach learned high fitness tanks in fewer generations than the competitive approach, but is slower due to increased amount of tank evaluations per generation. A parallel implementation can reduce the evaluations per generation significantly but will still be only as fast as its competitive counterpart. A comparison of the dominance hierarchies suggests that hybrid coevolution on average yielded higher quality solutions but by only a small margin. Despite the fact that no significant improvement was seen, this research has shown that the hybrid approach works well.