This is an AI that plays Connect Four using MiniMax with alpha-beta pruning, as well as accompanying programs that use the algorithms for the AI.

As such, there are 3 programs that can be built:

• Genetic algorithm in the ga directory
• Graphical game in the sdl-game directory
• Text-based game in the text-game directory

Binaries are in the Downloads tab above.

All programs require Go to compile. sdl-game also requires Go-SDL.

1. Change into the directory for the program (e.g. cd ga).
2. Run go build. This should give you an executable with the same name as the directory.

On starting, if the population file is specified and exists, ga will load the population, allowing for the resumption of running ga from a previous instance. If not, the population will be randomly generated from a uniform distribution over [-1,1]^6.

Everytime after a new generation is crossed over and mutated, the population is saved, along with the generation number, the best genome from the previous generation, and the fitness of that genome.

You start as the first player, red, while the computer plays the second, black. The board is shown as follows:

   B   
   R   
   RR  
   BB  
   RB  
0123456

It is red's turn.
Enter the column to place your piece:

R represents your pieces, while B represents the computer's pieces.

On each move, you enter the number of the column where you would like to place a piece, as shown on the bottom of the board.

You start as the first player, red, while the computer plays the second, black. Click on a column to place a piece.

The static evaluator function is essentially copied from Jenny Lam's report "Heuristics in the game of Connect-K." However, some slight changes were made. The resulting function is a linear combination of the following:

1. p_1 win: This is 1 when the AI is at a winning state, and 0 otherwise.
2. p_2 win: This is 1 when the oppenent is at a winning state, and 0 otherwise.
3. p_1 odd threats: number of lines of three of p_1's pieces that have an empty spot on one side, where the empty spot is in an odd row.
4. p_2 odd threats: likewise, but for p_2 instead of p_1
5. p_1 even threats: number of lines of three of p_1's pieces that have an empty spot on one side, where the empty spot is in an even row.
6. p_2 even threats: likewise, but for p_2 instead of p_1

where p_1 is the AI and p_2 is the opponent.

The coefficients for each of these is then found using a genetic algorithm.

Fitness is determined by counting the number of wins after running 5 rounds of trials, wherein each genome is set against another random genome. To prevent genomes from participating in a disproportionate number of trials, the second player is actually selected, in order, from a random permutation of all genomes.

For the genetic steps, roulette selection with uniform crossover is used. This is followed with a mutation operator, where a random number from the distribution N(0, 0.0009) is added to each gene.

After running this for 10 generations on a population of 100, the following coefficents were found:

( 0.2502943943301069,
 -0.4952316649483701,
  0.3932539700819625,
 -0.2742452616759889,
  0.4746881137884282,
  0.2091091127191147)

The order of these correspond to the values in the list above.

I cannot actually beat this static evaluator with these coefficients. However, I will admit that I am terrible at Connect Four, as can be seen in this video.

After further testing, an interesting preference was found. It will frequently ignore immediate wins in favor of setting up no win situations. This is not surprising given that setting up threats is a higher priority than winning. However, these setups do not always turn into wins. However, some method of supplying non-linear effects, as well as taking the depth of the evaluation into account to allow for uncertainty seems like a good idea.