Decision Transformer Model and Its Practical Use Cases

Abstract

This bachelor thesis investigates the Decision Transformer (DT) model, a recent method that reformulates reinforcement learning (RL) as a sequence modeling problem. Unlike traditional RL approaches that rely on value functions or policy gradients, DT predicts future actions based on full trajectories of states, actions, and desired returns (return-to-go). The thesis first presents the theoretical foundations of reinforcement learning and Transformer architectures, followed by a detailed examination of the DT model’s components, architecture, and advantages. Practical application areas are then proposed and evaluated through experiments in a grid-based navigation task and the game 2048. The results highlight the model’s strengths and weaknesses, and offer insights for future research and deployment in real-world settings.