This paper explores the use of the NEAT algorithm combined with self-play in deterministic game environments. Our objective is to investigate how quickly models can adopt specific strategies with minimal reliance on training data and expert knowledge. We developed a game similar to the board game Catan as a platform for training models. Multilayer feedforward neural networks were used to evaluate board positions, and games were played using a minimax search strategy. In each generation, neural networks competed against the best networks from the previous generation. Poorly performing networks were eliminated, and the survivors produced offspring through crossover and mutation.