Skip to content

Latest commit

 

History

History
124 lines (105 loc) · 3.67 KB

README.md

File metadata and controls

124 lines (105 loc) · 3.67 KB

ttt

An implementation of tic-tac-toe in C, featuring an AI powered by the negamax algorithm, the reinforcement learning (RL) algorithm, the Monte Carlo tree search (MCTS) algorithm. And the RL algorithm contains the Monte Carlo learning algorithm and TD learning algorithm.

Build

negamax AI

If you want to play with negamax AI: Build the game:

$ make

and run:

$ ./ttt

reinforcement learning AI

If you want to play with RL AI: Train the state value table first, you could modify the hyperparameters, which are macros in train.c:

  • MONTE_CARLO : Decide whether or not to use the Monte Carlo method or the TD method.
  • REWARD_TRADEOFF : Decide the reward of the Markov decision model, which is a balance between the episode reward and the score from get_score. Those who want to fully customize the reward can modify it in the init_agent function.
  • INITIAL_MULTIPLIER : Decide the initial state value, which is the multiplier of get_score in the initial state value. Those who want to fully customize the initial state value can modify it in the init_agent function.
  • EPSILON_GREEDY : Decide whether or not to use epsilon-greedy exploration when training.
  • NUM_EPISODE : the number of game episode in training
  • LEARNING_RATE, GAMMA : $\alpha$, $\gamma$ in training.
  • EPSILON_START EPSILON_END : $\epsilon$ in Epsilon-Greedy Algorithm and it would decay exponentially.

compile

$ make train

and run:

$./train

Build the game playing with RL agent, it would load the pretrained model from train:

$ make rl

and run:

$ ./rl

MCTS AI

If you want to play with MCTS AI:
There are several hyperparameters you can modify:

  • EXPLORATION_FACTOR in agents/mcts.h : The exploration parameter.
  • ITERATIONS in agents/mcts.h : Number of simulations in MCTS.

Build the game:

$ make mcts

and run:

$ ./mcts

Run

These program operate entirely in the terminal environment. Below is its appearance as it awaits your next move:

 1 |  ×
 2 |     ○
 3 |
---+----------
      A  B  C
>

To execute a move, enter [column][row]. For example:

> a3

Press Ctrl-C to exit.

Game Rules

The winner is determined by the first player who successfully places GOAL of their marks in a row, whether it is vertically, horizontally, or diagonally, regardless of the board size.

Using the following 4x4 board games, whose GOAL is 3, as examples:

 1 |  ×  ×
 2 |     ○  ×
 3 |     ○
 4 |     ○
---+------------
      A  B  C  D
>

The player "○" wins the game since he placed his marks in a row vertically (B2-B3-B4).

 1 |  ×  ×  ○
 2 |  ×  ○  
 3 |  ○  
 4 |     
---+------------
      A  B  C  D
>

The player "○" wins the game since he placed his marks in a row diagonally (A3-B2-C1).

 1 |  o  x  
 2 |  o  x  
 3 |  o     x
 4 |  o  x
---+------------
      A  B  C  D
>

The player "○" wins the game if ALLOW_EXCEED is 1; otherwise, the game will continue because the number of "○"s in a row exceeds GOAL.

Reference