jester is an attempt at creating a strong but potentially unconventional AI for the card game Durak.
jester implements a simpler variation of traditional Durak. This has the effect of simplfying the game's implementation and search space, with improvements in bot strength and less potential for bugs. The differences are as follows:
- Only one attacker and defender are allowed at a time. Other players are not allowed to join in an attack.
- Players cannot divert an attack by playing a card of the same rank (e.g. they cannot pass the buck).
- Only one card can be played at a time.
All gameplay is currently conducted through the console. There are no plans to add a GUI.
Players in jester behave according to several pre-defined options. When running the game engine, you can choose which players to instantiate in your games.
jester implements the following player types:
- Minimal — The minimal player will only play a card on the on the first attack. On defense, they will always yield. A minimal player represents the worst possible strategy and will always lose, even against random play.
- Random — If it is a random player's turn, they will select a valid, random action with no bias.
- Greedy — A greedy player is similar to a random player. However, a greedy player will always play a random card if it can.
- Determinized MCTS — A player implementing Monte Carlo tree search with explicit determinzations. That is, the player will fill in incomplete game information with several potential determinizations and then run MCTS on each state.
- ISMCTS — A player implementing information-set Monte Carlo tree search. Determinizations occur every simulation and all paths taken are merged into a single MCTS tree.
- Human — A human player interfaces with the console and is controlled through stdin.
Work-in-progress options are:
- Tabulated CFR — Pure counterfactual regret minimization. The algorithm is implemented as listed on page 70 of Richard Gibson's PhD dissertation.
CFR is a reinforcement learning algorithm that learns from playing repeated games against itself. However, the complexity of Durak, even with 36 cards and a simplified rule set, means that CFR needs a large amount of computational resources in order to raise its playing strength to an acceptable level. For reference, NLTH poker bots using CFR are usually trained on supercomputers, where millions of CPU hours are invested into training. For this project, that may not be feasible.
That being said, CFR can quickly be trained to play optimally on an extremely
simple version of Durak involving only 8 cards and 2 players, requiring only
a few minutes of multi-threaded training. You can use this reduced form
of Durak by providing the -r flag when running the executable. Training
mode can be enabled via the -t flag.
At the moment, the results of CFR training are saved to a Redis database.
You can specify which database CFR should connect to using the REDIS_URL
and REDIS_PORT environment variables. Otherwise, REDIS_URL will default
to localhost and REDIS_PORT will default to 6379.