Human-Aligned Chess AI: A Multitask Transformer for Humanlike Decision-Making

At the moment, it's clear that AI has surpassed human ability in almost every field. But, how useful really is this to us? Several fields (eg. law, education, games) have noticed that having a “perfect” AI isn't as good as it seems in all cases. One of the marquee examples is law, where an AI wouldn't consider highly important parts of human thinking such as fairness and ethics, it would sentence the victim based on raw logic, something that we clearly want to avoid. Being able to make AI that can understand humans and make decisions like them is becoming much more important in order to be able to make critical decisions that others can interpret and collaborate with. Being able to completely model human reasoning in any field is not possible due to the amount of ‘noise’ in each individual, but we aim for this in the game of chess. Chess is one of the most complex and popular games due to its large fan base and the vast amount of positions that can arise, despite the test of time chess remains unsolved. There are multiple existing chess engines, such as Stockfish and Alphazero, which are essentially ‘perfect’ forms of AI, they aim to be as strong as possible and win as many games as possible rather than trying to play like a human. We introduce Harmonia, an encoder-only Transformer trained through multitask learning to predict human chess moves and game outcomes. Although there have been multiple attempts to emulate human play in chess through the Maia project, Allie Chess, etc., our model is able to surpass all of these existing baselines despite being significantly smaller in size and trained on fewer data. Although Harmonia is extremely good at modeling human play in chess, it is much more important to promote the development of human-AI interaction. It is unique in the sense that it was trained on human data in order to learn human policies and internal representations, which can be used as a tool for further research on educational uses and understanding how humans think. The results in this paper show that there is a high potential for artificial intelligence that can mimic human behavior to be more understandable and collaborative with fellow humans.