Limitations and Future Work

This work has not used the method of Minimum Message Length (MML) [10][11] to assess the predictive ability of the strategies inferred and this is a possibility that will most likely improve the predictive ability of this method, as this does not take into account any prior beliefs. The chunking approach used by other researchers [34][33][15][24] could be adopted here. The process would require finding geometric patterns in the chess pieces and then placing them into certain categories. These classes might then be placed into a decision tree (or maybe, more than one) that can be used for classifying a position; starting from a generic form and becoming more specific. In the leaves, the approach used in this paper could be used. Therefore, for each leaf a different evaluation function and search style could be inferred. This would mimic humans, as they look at certain configurations and then make a judgment as to what is considered important and the style of searching required to make a move. This method would then be able to implicitly differentiate between the opening, middle and end games and play the appropriate strategies, as needed.

A further improvement would be to increase the number of strategies compared and introduce methods that are more exotic such as conspiracy search [23]. Also calculating the Kullback-Leibler distance between the true model and the inferred model would determine the accuracy of the models. Another expansion that may be possible is probabilistically identifying the player of a previously unseen game.