A Combined Position Evaluation Function in Chinese Chess Computer Game

In Chinese-chess computer game (CCCG), the position evaluation function plays a crucial role in building a game playing program. Traditionally, there are two typical types of evaluation functions: standard heuristic evaluation function (SHEF) and self learning evaluation function (SLEF). The SHEF depends on the board position features to large extent, but it hardly includes all the features due to the limit of knowledge of the designer. The SLEF can explore the knowledge hidden in the current position which is difficult to find in the SHEF. In this paper, a combined position evaluation function (CPEF) is designed by weighted sum of SHEF and SLEF. SHEF considers the material balance and adjunctive value of position while SLEF takes the form of a three-layer fully-connected feed forward neural network. We use temporal difference learning (TDL) to train the neural network on professional game records. Based on the combined position evaluation function, a Chinese chess program HBUCHESS is developed. We experimentally validate that our CPEF is quite effective through competing with different kinds of testing players. With the help of CPEF, the intelligent level of HBUCHESS can be improved incrementally with the increase of number of professional game records SLEF learned. Furthermore, in the process of learning professional game records, we find that the performance of HBUCHESS is mainly relevant to the following four aspects: (1) the initial heuristic knowledge, (2) the number of nodes in hidden layer of neural network, (3) the trace decay parameter λ, and (4) the learning rate α.