Automatic generation of multi-objective neural game controllers using Pareto-based differential evolution

• Main Authors: Chin, Kim On, Jason Teo, Chang, Kee Tong
• Format: Research Report
• Language: English
• Published: Universiti Malaysia Sabah 2010
• Subjects: QA Mathematics
• Online Access: https://eprints.ums.edu.my/id/eprint/22850/1/Automatic%20generation%20of%20multi-objective%20neural%20game%20controllers%20using%20Pareto.pdf; https://eprints.ums.edu.my/id/eprint/22850/

Eversince the born of personal computer to this world, computer games have not been stoping in their development. More and more of computer games are created and more and more genre exsist. RTS games generally are war simulation. The game features of RTS game are very similar to a real war where player will have to manage their resource income while planning their army, technology, attack location, attack time and much more. Player need to be smart in managing and planning their strategy to overcome their enemies. This also means that RTS game can be a very powerful! testbed for AI. Many researches have done their reseach on improving RTS games such as improving the micro and macro management of the game AI. This report is to investigate how well does Pareto-base Differential Evolution (POE) with a Neural controller perform in a RTS game. A popular billion dollar RTS game, Warcraft III is chosen to be the testbed. For this research, three Evolutionary Algorithm (EA) (Genetic Algorithm (GA), Differential Evolution (DE), and POE) is use to evolve a Feed-Forward Artifitial Neural Networks (FFANN) to playa custom made map in Warcraft III and the outcome is compared amng them. Resutls clearly shows that the hybridized ANN can defenately generate a controller to play the game. DE with FFANN shows a better results comparing to other two approaches. However, PDE with FFANN approach explored a much cheaper of army to win the match.