It’s beating me so far. I’m not going to continue playing – more of a proof of concept. It’s keeping track of the scoresheets and it has a mechanism to pick the dice you want to hang on to.
So, that’s one game an AI can play without any problems.
Since this is FQ: AlphaGo (the specific instance of AlphaZero used for Go) was given expert games as part of its training; the AlphaZero instance of chess was not. Both models also used self-play.
If there is an online version of a game available, it would be a lot easier to take screenshots of the game progress than photos of an actual gameboard. I think i will give it a shot with Gemini later this eve. Game TBD, maybe Risk! That would be fun. Or Scrabble. I’m interested to see if it would start making up words based on the tiles that it has.
Here’s a Catan Clone called “Colonists” that’s free and online, with rules available.
I think it’s obvious I know very little about ai’s.
I think Chess or Go would be considered fairly medium weight games in modern board gaming. They just have a sense of historical mystique.
But Chess and Go aren’t really any more complex than games like Onitama or Hive and nobody is calling those heavy weight games.
I feel an ai would have a lot more difficulty in a game like Arkwright or Feudum or Hegemony.
MuZero is a successor to AlphaZero that learns to play games without rules. AlphaZero learned to play chess, shogi, and go through self-play.
MuZero is a computer program developed by artificial intelligence research company DeepMind, a subsidiary of Google, to master games without knowing their rules and underlying dynamics.
MuZero was trained via self-play, with no access to rules, opening books, or endgame tablebases. The trained algorithm used the same convolutional and residual architecture as AlphaZero, but with 20 percent fewer computation steps per node in the search tree.
I’ve been working on this as a fun project with the goal of creating a tool to prototype board games and mechanics. Rather than giving it images of the game or a PDF of the rules, I have a game engine that plays the game and enforces the rules. The Player or Bot just gets a set of valid choices and the game’s state.
The dumbest Bot picks a random choice. My intention was to use Reinforcement Learning to create an ML/AI Bot, but I haven’t gotten to that part. Recently I have been trying to use an LLM to generate code for the Bot with (so far) not so great results.
There are AIs that can play Atari games using just the screen contents. In one of Andrej Karpathy’s Standford class sessions he creates a simple AI to play Atari Breakout. It is a pretty short implementation and you can follow along and build your own. So going from valid choices to a picture of the board game is possible, but more computational expensive.
My kids and I were discussing this topic recently. Abstract board games, like Chess and Go, might be medium-weight in rules, but not necessarily in optimal game play. I assume Onitama and Hive as well, but I am not familiar with them.
Heavy-weight, non-abstract games might have more choices and complexity, but they are more prone to being less balanced. There can be decision paths that are clearly better or worse and an AI can learn these paths quickly.
As an experienced chess coach, I can say two things:
- it’s easy to teach ‘how to play chess’ … there are only 6 pieces, plus a few special moves (e.g. castling)
- chess strategy and tactics make it a deep complicated game to do well at … however since it’s a logical, full information game, it’s ideal for computers to pick up and excel at
This is part of the discussion that needs to be explicit - the rules are “what is a valid, legal move?” and “how does one win?”. Strategy is what makes a player (real or AI) good. Part of strategy is understanding what an opponent could do in response.
I assume with the “not told the rules” version of AI, there is still a supervisor validating any moves. Otherwise the result would just be decorating the chessboard randomly. So it is being told the rules by inference. After a thousand games, an AI might enocunter and figure out all the rules and limitations of, say, castling. I presume the supervisor does not say “you cannot castle through check” or any other hint, simply says “move not allowed”.
That is my assumption as well. In chess engines there are various formats for encoding the legal moves at any given step. I think this is the input to the AI and the output is one of those moves.
I can see this process of learning by self-play working in chess, where you have thirty-two pieces on a sixty-four square board and a single winning objective. A fast computer can play the game millions of times and develop paths that will lead to winning.
But would this process work in a game like Bridge, for example? Suppose an ai was “playing” a hand of thirteen cards, with no programmed idea what the rules are. So it plays a random card. The other players, human or ai, play cards as well and then the judge informs the table who won the trick.
An ai playing the same hand over and over again would eventually learn which cards win the trick and which cards lose. If you had a really long time, it would eventually learn what order to play all thirteen cards in to produce the highest total of wins.
(To give an idea how long this would take, each player has thirteen cards, which means there are over six billion different ways those cards can be played. And to determine which way is optimal, you would have to compare each of those six billion ways to each of the six billion ways the other three players could possibly play their hands. Over 1,300,000,000,000,000,000,000,000,000,000,000,000,000,000 plays if my quick calculations are correct.)
But at the end of these tredecillion plays, the ai will have learned the optimal way to play this hand.
And will be ready for the new deal, which will be completely different.
I don’t have an opinion on whether bridge is workable, I’d just note that every hand opens with bidding in which your bid is often communicating things to your partner that may sound like that bid at all (i.e. conventions), I’m not sure how that part would work. But I’d love if it came up with new conventions somehow.
ETA also, the bidding of your opponents can inform the optimal play of your hand. So there’s that as well.
I think the trading aspect of Catan might be the hardest part.
Wonder if you started with no luck (other than initial set up) games like Caylus (initial turn order, and order of pink buildings) or Ora & Labora (initial turn order, and whether it is Ireland or France) would be easier. Puerto Rico has a smidge of randomness (which plantations come out)
Brian