A methodology for revealing and monitoring the strategies played by neural networks in mind games

An effective approach in the application of computational intelligence to mind-games is neuro-evolution. Neural networks are efficient at autonomous learning and pattern recognition tasks, which has led to many outstanding accomplishments in mind-games' playing ability. Neuro-evolution is a practical way that uses an evolutionary framework to train a neural network in a complicated context where credit assignment is hard. However, the neuro-evolution process may stagnate, or result in solutions with a limited quality. Potentially a cause for this problem are the limitations in understanding the neuro-activities as evolution progresses. A possible solution lies in unfolding the dynamics of the evolution process and the types of the strategies evolved as the evolution progresses; thus providing a diagnostic tool in real-time to identify neuro-dynamic causes of stagnation. Rule-extraction techniques are a notable solution to understanding the networks evolved. However, the extracted rules lack the necessary expressiveness to explain game-playing strategies. We call these rules a syntactic representation of the network that lacks semantic power. In this paper, we will present a methodology whereby a computational environment is used to unfold the evolution of a mind-game neuro-player; thus providing semantic power. Within this environment, we propose to extend the role of computer players to act as a “cognitive” functionality model, thus providing deeper kinds of explanations. We use the game of Go to demonstrate the functionality of the methodology. We then demonstrate that this methodology is successful in determining the types of strategies evolved in a neural Go player, and in monitoring the dynamics of the evolution.