“First, solve the problem. Then, write the code”
The purpose of this section is to lay down my thoughts and ideas while I take a journey into the world of Mancala and begin to develop an algorithm to evolve an agent to play an African board game of counting and strategy.
The approach which I will use to evolve the game agent is competitive coevolution to evolve the evaluation function to be used in an alpha-beta game tree. The evaluation function will be in the form of a Neural Network.
Since I will not have a training set of board states and target moves, training will need to be unsupervised. I will use a Particle Swarm Optimizer to train the neural network in a competitive coevolutionary manner. The learning model will consist of three components:
- An alpha-beta game tree expanded to a given ply depth (Can thus simmulate difficulty level). The root of the tree represents the current…
View original post 269 more words