TicTacToe game with Double Q-Learning
Aim of this project to build a model-free reinforcement learning algorithm (QLearning) that can play tic tac toe better than a human does. When application runs, agent starts playing agains in two other agents. One opponent agent picks highly random moves while the other one makes a bit smarter moves.
At the same time you can play against the agent, by using numbers on the board between [0, 8]. 0 refers to the (1, 1) cell in the tictactoe board while 8 refers to (3, 3).
Building
> cmake . && make
This command should create a binary in bin/ folder name game.
Running the game
After running the command from the base (ttt_qlearning) folder,
> bin/game
You will generate 3 agents playing TicTacToe.
- Agent A will be playing with Agent B.
- Agent A will be playing with Agent C.
And you will be promted to enter a number to mark on the board. While training you can play against Agent A yourself and see the improvement.
[agent a vs. b] agent 0 won %: 58.547, agent 1 won %: 41.0874 | agent 0 double table (action) size: 29936
[agent a vs. c] agent 0 won %: 51.6226, agent 1 won %: 48.2668 | agent 0 double table (action) size: 30034
-o-
xo-
x--
Enter pos [0-8]:
While training continues you can play the game continuously.