The goal of this work was to help the diagnosis of MCI with games. From many possibilities I chose five-in-a-row (gomoku). Because of the game’s complexity I developed agents which are capable to evaluate the efficiency of a move in a possible situation. This allows me to use the agents as opponents for a human player or to use it to evaluate a game as an observer.
The non-learning agent is entirely algorithmic. I used it later to teach other agents. My goal was not to create a perfect player but rather a strong enough opponent.
One learning algorithm used reinforced learning with pattern recognition. I researched different reinforcing methods like self-learning and learning by observing two skilled players. The agent contains its own database equipment, and data-compression method. I tested its capacity, performance, storage ability and the limits of the agent.
The other learning agent is a neural network. For the network I chose MLP. During the teaching I observed the different effects of changing the structure of the net and the input database.
The agents were tested on a control group and I compared the results with the players’ actual performance.