An Empirical Analysis of Tabular Q-Learning for Gridworld Navigation

This paper presents a controlled empirical analysis of tabular Q-learning is presented for a Gridworld navigation task modelled as a Markov decision process (MDP) with obstacles and a reward-bearing jump tile that teleports the agent. Exploration is handled using a decaying ε -greedy policy. Learning is evaluated under a fixed budget of 100 episodes with performance-based early stopping, and results are reported over multiple random seeds. Sensitivity to key hyperparameters (learning rate α and discount factor γ ) is analysed, and the learned state-value function and greedy policy are visualised to support interpretability. Across the evaluated settings, the learned policies reliably reach the goal, and performance trends are reproducible across seeds.