Dream v3 z Kolmogorov-Arnold Networks

The goal of the project is to evaluate the potential for integrating a neural network architecture based on the Kolmogorov-Arnold theorem (KAN) with the model-based reinforcement learning (model-based RL) paradigm, as represented by the Dreamer v3 agent. The proposed methodology involves creating dedicated KAN implementations that can be used in place of conventional multilayer perceptrons (MLPs) in the fundamental decision-making components—the Actor and Critic networks. The main research hypothesis is that the superior approximation properties of KAN will translate into a measurable increase in sample efficiency and the agent’s ability to generalize in the context of complex, dynamic decision-making processes. The performance of the modified agent will be rigorously compared with the original implementation on two key benchmarks: Atari 100k, which evaluates learning speed with limited data, and the challenging Minecraft environment, which tests hierarchical planning and adaptation capabilities.

Machine-translated
Przetłumaczone maszynowo