Bridging Genetic Algorithms and Gradient-Based Learning: A Case Study on the Dinosaur Game

This paper proposes a hybrid approach that integrates Genetic Algorithms (GA) with derivative-based training for Multi-Layer Perceptron (MLP) neural networks in the Dinosaur Game environment. GA is employed to generate training data in the absence of existing datasets, and the performance of derivative-based MLP models trained on this data is evaluated. The performances of the various network architectures have been compared with respect to network structure and activation functions. The generalization capability of the network architectures has been assessed on tampered test environment.