Table 21.

Ablation study results for DQN showing the impact of removing individual simulation components.

Ablation setting	Mean reward	Relative change	Interpretation
Full model	Best	—	Rich dynamics allow DQN to exploit contextual, stochastic, and long-horizon structure
No rich state-action effects	↓↓	Large decrease	Removes context-dependent leverage, collapsing advantage of deep function approximation
No adherence stochasticity	↓	Moderate decrease	Eliminates uncertainty signals critical for robust policy learning
No behavioral drift	↑ (artificial)	Inflated	Unrealistically optimistic rewards due to stationary behavior
No dropout/re-engagement	↑ (artificial)	Inflated	Removes long-term engagement penalties and overestimates policy quality
No environmental modulation	↓	Moderate decrease	Reduces contextual diversity, limiting state generalization
No misreporting noise	↓	Small decrease	Simplifies perception, reducing robustness but less critical than dynamics

Performance degradation is reported relative to the full model.