Game kernel design using genetic programming

Abstract

In the recent game developing trend, one of the most important parts is Artificial Intelligence (A.I). To be specific, it is about the controlling part of Non-player character (NPC). Basically, the A.I can be divided into 2 types: 1) hard-coded A.I, 2) A.I with learning ability. The first type of A.I – hard-coded A.I only gives the particular response based on different conditions. On the other hand, the second type of A.I - A.I with learning ability has self-improve feature that can learn and evolve to the correct strategy. This project, Game Kernel Design Using Genetic Programming, aims to implement a learning system to the A.I that the A.I can learn the correct strategies from players. Genetic Programming (G.P) was used on a Tank fighting game to generate an A.I that gives response which is similar to human beings. In this G.P system, Binary-tree representation (BTR) and offline update training method were used for generating the A.I of the NPC. To simulate the player, data of players’ battle was recorded as the offline training set. Moreover, the human-like action becomes the action set of the A.I. For the evolution, the tournament method was used as the selection method in each generation. A.I. would have a prominent improvement after 200 generations. Compared to those randomly generated strategy, the improved A.I. would simulate human strategies. Therefore, it can easily defeat any of them.