Evolving a designer-balanced neural network for Ms PacMan

Balancing games towards designer requirements is an on-going research area with proven potential for use in industry. However, other elements beyond game mechanics can be tweaked and optimised to offer a rewarding gaming experience to players. This work looks at using proven techniques and tools to change not the parameters of a game, but the parameters of an agent playing a game to create a version of that agent that behaves in a designer-specified manner. These new agents can then be utilised for a wide variety of tasks, from offering new challenges to players, to aiding designers in automating parts of their pipeline when balancing games towards given requirements. Through the use of genetic algorithms and intelligent collection of game metrics, we are able to successfully generate varying neural-network controlled agents, each with different styles of play and levels of skill.