An improved q-learning algorithm with optimized initialization and annealed boltzmann exploration

Abstract

This paper proposes a hybrid enhancement method for Q-learning that combines direction-sensitive Q-table initialization with annealing-based Boltzmann exploration. Initialization leverages geometric priors to bias actions toward the target without leaking obstacle information; the annealing-based Boltzmann method achieves a smooth transition from extensive exploration to exploitation. By leveraging the symmetry of isometric states and an adaptive exploration strategy, the improved Q-learning algorithm achieves faster convergence in discrete action environments.