rummikub

This package aims to provide a Rummikub implementation with a machine learning flavor

Evaluation function

Example of search function for board game: Expectiminimax

Deep Reinforcement learning

cf: https://github.com/deeplearning4j/rl4j

Domain

MDP 5-tuple (S, A, T, γ, R.):

– S = {s1, s2,...} is the possibly infinite set of states the environment can be in. All the tiles on the tables in the rack? – A = {a1, a2,...} is the possibly infinite set of actions the agent can take. Tile allowed at any given point in the game. Basically the tiles that can be played on the table? – T(s|s, a) defines the probability of ending up in environment state after taking action a in state s. Depends on other players decision and random picks from the pool? – γ ∈ [0, 1] is the discount factor, which defines how important future rewards are. The value of the tile that I can put on the table ? – R(s, a, s1) is the possibly stochastic reward given for a state transition from s to s1 through taking action a. It defines the goal of an agent interacting with the MDP, as it indicates the immediate quality of what the agent is doing.

Reinforcement Learning Algorithms

Look at [5] $4

Q-Learning

In our case Q(s, a) expected value of taking action a in state s.

References

• [1] Abstracting Reusable Cases from Reinforcement Learning
• [2] A brief tutorial on reinforcement learning: The game of Chung Toi
• [3] Reinforcement Learning for Board Games: The Temporal Difference Algorithm
• [4] Multi-Stage Temporal Difference Learning for 2048-like Games
• [5] A Gentle Introduction to Reinforcement Learning
• [6] AI Mahjong
• [7] DominAI
• [8] Giraffe