This thesis examines the application of Decision Transformers (DTs) in developing affect-based agents that mimic human player arousal in a video game environment. By training DTs on six offline reinforcement learning datasets, generated via Proximal Policy Optimization (PPO) and Go-Explore under reward schemes prioritizing game score, arousal, or a blend of both, we demonstrate that the DT architecture can successfully learn arousal-driven behaviors. The resulting DT agents not only replicated the intended affective patterns but also consistently outperformed the original agents used for data generation. The results underscore that dataset diversity, particularly from the Go-Explore method, is crucial for balancing game performance and affective emulation, though overfitting remains a significant challenge. Ultimately, this research validates the effectiveness of sequence modeling for creating more nuanced, human-like AI agents and provides key insights into how offline data generation strategies shape their behavior.