An Artificial Intelligence Player for the Game of Inverse Chess

This paper describes the rule base and a computer player architecture for a new 2-player combinatorial game related to chess, called the Game of Inverse Chess (IChess). This game can be summarized as playing chess backwards in time, from any end board state to the initial state, in a competitive fashion, with players alternating turns. The first player to achieve the starting configuration (or other predetermined winning configuration) is the winner. The moves in IChess are generally the moves in chess played backwards (e.g. bringing in a piece into the board (spawning) instead of killing, king giving an inverse check...). There are a couple of exceptions pawns can be brought in and also removed. The game is quite enjoyable, with about 70% of a few hundred college students liking it in trials. It is also computationally very challenging, from multiple viewpoints game theoretic, graph theoretic, etc. The state space of IChess is larger than that of Chess, since there are non-chess positions reachable in IChess. As such sophisticated AI based approaches are needed to develop a competent computer player. Our AI player uses (pruned) min-max search, plus heuristics which try to emulate the logic of a human player. The min-max search is based on a board evaluation metric (BEM), which is a non-linear function of board state features. The BEM itself can be improved using learning techniques. Our AI player is capable of beating a beginner level human opponent when run on a contemporary laptop (3GHz, 512 MB RAM unoptimized). We expect to be able to present results from a player which is capable of beating about 90% of human opponents by the time of the conference.