Skip to content

Latest commit

 

History

History
44 lines (31 loc) · 1.56 KB

README.md

File metadata and controls

44 lines (31 loc) · 1.56 KB

2048 Artificial intelligence

Description

What ?

The goal is to make an artificial intelligence on the game 2048.

Why ?

Because it's fun.

How ?

For the moment, there are only few algorithms written:

  • minimax
  • alphabeta
  • expectiminimax
  • 2 different genetic algorithms

Choice of the programming language

Obviously Python because it's easy to learn, to read, to debug. In addition, I'm using:

  • pygame to manage the 2D graphic interface of the game.
  • numpy to manage array calculations in the neural networks
  • matplotlib to draw and show graphs

There are easy to use.

Algorithms

Two different versions of genetic algorithms (file genetic_algorithms.py):

  • GeneticAlgorithm1: The first generation is composed of artificial intelligences (A.I.) with random neural networks weights. After they have played, best A.I are selected. A new generation is created depending on 3 groups: a group with the best A.I., a group with A.I. randomly crossed with the best A.I. and a group with the best A.I. but some of their genes are mutated. All A.I. are either crossed either mutated
  • GeneticAlgorithm2: it almost the same but among all A.I. to cross or to mutate, some ones won't be neither crossed neither mutated.

In the file decisions_algorithms.py:

  • algorithm minimax
  • algorithm alphabeta : more optimized than minimax
  • algorithm expectiminimax : more efficient than minimax

Results

The convergence of genetic algorithms are slow. More tests must be done to get a better configuration. Decision algorithms are efficient. The best one is expectiminimax.