Occasionally I google around looking for the latest research into hearts AI, and I never find anything. As far as I can tell the best hearts AI out there is still far below the level of an expert human.

I'm wondering if anybody knows of any recent papers or results. I'm particularly curious if a Monte Carlo Tree Search based approach combined with deep learning and reinforcement learning has been attempted.

(Please let me know if there is a more appropriate Stack Exchange community for this question.)

  • 2
    Back in 1998, Windows Hearts creamed me. Also I was 11, that may have been a factor. – corsiKa Jul 8 '16 at 15:51

Update based on your comment: I gather that you are more interested in research than making a game. In that case I'll point you to a couple of papers that might be helpful. This paper talks about what makes Go a unique challenge, and the introduction has a quote (and some references) you might find useful:

Minimax search variants are particu- larly effective in Chess, Checkers, Othello and Backgammon, whereas Monte-Carlo simulation has proven most successful in Scrabble and Hearts.

Regarding hearts, the reference is to this paper which is unfortunately not free but maybe you have access through an employer or university.

I hope that helps. I still recommend the programmers.SE question I linked below; there's some good discussion there about Monte-Carlo methods as they relate to card games.

State of the art? Maybe Deep Mind. It's a general purpose AI that seems to excel at a great many games. Not sure if it's been applied to hearts, but I don't see why it couldn't be.

Realistically? Maybe Monte Carlo. You can read some discussion of it here: https://softwareengineering.stackexchange.com/questions/213870/best-techniques-for-an-ai-of-a-card-game

If you want to make a Hearts game, as opposed to an AI research topic, remember this: Games are supposed to be fun. An AI that could beat practically any human player wouldn't be fun for practically any human players. Fun AI is what Søren Johnson focused on for Civilization IV, and he has some good thoughts on it in this video. Furthermore, it's usually easier to develop an AI that's kinda smart that one that's really really smart.

| improve this answer | |
  • 1
    I highly doubt Deep Mind's Atari player (which maybe you are referring to by "general purpose AI") will be of any use in playing a game like hearts. I'm sure their Alphago system could be adapted to play hearts, although there are some key differences: hearts is multiplayer and has imperfect information. Those differences seem significant enough to make for interesting academic research. Just wondering if anybody has attempted it. – dshin Jul 12 '16 at 19:39
  • In the past several years, there have been several efforts in applying the ideas of AlphaGo's architecture towards games involving multiple players and imperfect information. DeepMind created AlphaStar to play Starcraft, OpenAI created Five to play Dota, and the University of Alberta created DeepStack to play poker. It seems to me the same approach will work in Hearts. Hopefully some enterprising student will come along soon and create a superhuman Hearts agent for a Master's thesis. – dshin Aug 10 '19 at 1:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.