Skip to content

Latest commit

 

History

History
55 lines (38 loc) · 2.14 KB

README.md

File metadata and controls

55 lines (38 loc) · 2.14 KB
Raw

Dots_and_boxes

Dots and Boxes is a simple 2 player game, played on square grid. In this project we apply methods of reinforcement learning for training an automatic player. The method used is a version of Q-learnig adjusted for this particular game to allow smaller memory costs.

The motivation, details of the implementation and the results are summarized only in slovak language in the report latex/projekt-ML-dots-and-boxes.pdf

In the process of working on this project, several python programs were developed.

Programs

Running the $ python play.py in the terminal you can play against my best trained bot2(10) or you can easily modify this file to play against another automated player.

  • dots_and_boxes.py contains all the global variables (i.e. BOARD_SIZE) and functions that can be universally used for developing, training and testing new automatic players. It is imported in all other used python programs.

  • bot1.py contains simple heuristics for automatic players

  • bot2.py contains functions used for training, saving and using a powerfull player which plays last k moves optimally and the rest by simple heuristics always4never3 inpired by another Dots and Boxes automatic player

  • train1.py is not really working, it was intended to provide functions to train Q function as MLP neural network (or decision tree) but it had very poor performance while testing its parts so I gave up trying and didn't even finish it.

  • train2.py can be used to training and testing of the Q-learning automated player - just uncomment the bottom part and play with it. Or you can try to change the training process.

    • If trained Q is saved (recommended as json) it can be used in play.py by loading Q and using function get_Q_move_and_update to get moves in play.py on line 329. inserting
move,update = get_Q_move_and_update(Q,board,l_rate=0,random_move_prob=1)
selected_x, selected_y = move

instead of currend code for getting bot moves.