Emergent Strategy in Wordle via Reinforcement Learning

Advisor: Dr. Michael White Status: Proposal & Experimental Design Complete

Project Overview

This project establishes the theoretical framework for a Reinforcement Learning (RL) agent capable of playing Wordle. I have designed an experimental pipeline to test whether a Large Language Model can derive optimal information-theoretic strategies (such as entropy reduction) purely through interaction, without explicit instruction.

Methodology

• Agent Architecture: The design utilizes a Llama-based policy model optimized via Proximal Policy Optimization (PPO).
• Reward Engineering: I have formulated a custom reward function that penalizes inefficient guesses to strictly incentivize information maximization.
• Evaluation Metric: The proposal includes the design of a “Shannon Entropy Expert” solver to serve as a mathematical upper bound for benchmarking the agent’s emergent intelligence.

Research Hypothesis

I hypothesize that under this specific reward structure, the agent will naturally converge on advanced human-like tactics—specifically Trap-Set Analysis—demonstrating that linguistic models can ground reasoning in information-theoretic principles.

Technical Focus

• Concepts: Reinforcement Learning (RL), Information Theory, Experimental Design
• Tools: Python, PyTorch (Planned Implementation)