Wordle Duel: Can Deep Reinforcement Learning Architectures Outperform a Heuristic Entropy Maximizer?

This thesis provides a comparative analysis of Information Theory and Deep Reinforcement Learning Approaches to solving a popular word puzzle game: Wordle. The game challenges players to deduce a five-letter hidden word within six attempts, providing feedback for each guess indicating letter correctness and positioning. Its constrained action space, limited to valid five-letter words, combined with its stochastic nature and sequential decision-making, makes it an ideal test bed for evaluating and contrasting entropy-driven strategy, focused on uncertainty reduction, with adaptive Deep Reinforcement Learning's dynamic learning under uncertainty. The primary objective is to rigorously compare these methodologies, particularly examining their performance under the critical constraint of a restricted guess space, implemented by allowing the algorithms to guess only words consistent with the feedback pattern observed after each attempt. By forcing algorithms to guess only words consistent with the feedback, this approach mimics the intuitive strategy of "narrowing down" the possible hidden word, which is often used by casual players. Simultaneously, this restriction reduces computational complexity and allows for a dynamically limited action space where valid guesses are strictly confined to the set of words still possible given the clues received, making the comparison between all the tested approaches easier. Additionally, the target word is uniformly sampled from a predefined lexicon, eliminating real-world biases to isolate algorithmic performance under fair conditions. By enforcing these constraints, the analysis evaluates how Reinforcement Learning’s capacity for long-term reward optimization compares to the entropy method’s stepwise uncertainty reduction, with the goal of minimizing puzzle-solving attempts.