Machine-Learning-Tetris

A Machine Learning Algorithm that learns to play Tetris

How it works

Fitness Function

Uses a fitness function to compute the score of a board. The fitness computes the score by calculating various features of the board (like max height, number of holes etc). The fitness function receives a set of weights that determine how important each feature is. The final score is computed by multiplying a weight with its corresponding feature for each feature.

The algorithm runs the fitness function on all possible next boards, and does the move that has greatest fitness. (a next board is the current board, with the current piece added in any way)

Genetic Algorithm

Uses a genetic algorithm to fine-tune the weights that correspond to the features, to find the perfect set of weights that result in maximum height.

Mutated Weights

The best two weights are taken into the next generation.

First, coefficients from the two weights are randomly taken to create new weights

Then, random mutations are put in all weights. The amount by which a mutation changes the value is a function of the average score for a generation. Further, mutations which create small changes occur more often, while mutations that create large changes occur less frequently.

Understanding output

score0: .....
.
.
.
score0:   282 score1:   398 score2:   364 Gen: 9 W: -0.0590602  -0.288204  0.0293877   0.519238 AvgScore(of 3): 348
score0:  1270 score1:   676 score2:    64 Gen: 9 W: -0.0590602  -0.281186  0.0447034   0.676541 AvgScore(of 3): 670
score0:   101 score1:  2951 score2:   540 Gen: 9 W: -0.0590602  -0.333541  0.0293877   0.392932 AvgScore(of 3): 1197.33
--MaxOneAvgScore: 1596.33 MaxOneIndex: 15 MaxOneWeights -0.0591585  -0.288204  0.0367518   0.473326 --
--MaxTwoAvgScore: 1223.33 MaxTwoIndex: 7 MaxTwoWeights -0.0590602  -0.288204  0.0367518   0.473326 --
------- Gen: 9 MaxAvgWtScore: 1596.33 AvgGenScore: 513.253 --------

score0, score1, score2 are the scores from running the weights W: weight1, weight2, ... AvgScore(of 3) is the average of score0, score1, ...

MaxOneAvgScore is the maximum average of the list of weights. MaxOneIndex is the index of the weight being used, MaxOneWeights are the weights of the maximum average.

Gen is the generation number, MaxAvgWtScore is the maximumAverage of the generation, AvgGenScore is the average score of the weights

Command line arguments

./out to run

-gen numberOfGenerations (default is 100) (might take some time on default settings)

-wts numberOfWeightsPerGeneration (default is 30)

-gpw numberOfGamesPerWeight (default is 3)

-seed to set the seed (initial value is time(NULL))

-play to show the play of a single game

      The following are only applicable if -play has also been given:

      -coefficient to give your own coefficients (between -1, 1)

      -graphics adds graphics (not yet implemented)

Built With

• c++
• javascript