Extending Heuristic Knowledge Transfer for General Game Playing

General Game Playing (GGP) is a field of study in which artificial agents are required to compete in games whose rules are not known until runtime. Monte Carlo Tree Search (MCTS) is popular in this domain for its ability to quickly simulate many instances of a game and run in whatever time it is given. Heuristics may be used to guide these simulations, but since the game is not known in advance, a typical MCTS agent cannot know which heuristics are useful, and must expend precious time to discover them. However, an agent able to take advantage of prior knowledge from other games can do better. This paper extends our publication at the 2024 IEEE Conference on Games [1]. As in that paper, we present a technique for automatically transferring heuristic knowledge between distinct, but similar, games. We show that this leads to better performance in games within the GGP framework, especially when initialization time is short, and we show that negative transfer is possible to detect and avoid. Unique to this paper, we give experimental results for an alternate method of negative transfer protection, and explore the application of our methods to single-player games.