Enterprise AI Analysis

Action-Aligned Simulator Calibration

Traditional simulators optimize for average accuracy, often overlooking "decision-critical regions" where small errors can have large impacts. Sim2Act introduces an adversarial calibration method that reweighs surrogate outputs based on decision-critical errors. This ensures the simulator is most accurate where it matters most for action ranking and decision outcomes, preventing mispredictions from destabilizing policies.

Policy Robustness with Group-Relative Perturbations

To prevent policies from becoming overly conservative due to potential errors, Sim2Act employs group-relative perturbations. Instead of reacting to single noisy states, the policy learns to compare actions across a coherent group of perturbed states. This approach stabilizes relative action preferences, allowing the policy to distinguish between high-risk high-reward actions and high-risk low-reward actions, maintaining robustness without sacrificing performance.

Experimental Findings

Extensive experiments on supply chain benchmarks (DataCo, GlobalStore, OAS) demonstrate Sim2Act's superior performance:

• Consistently outperforms existing robustness baselines under various perturbation settings.
• Achieves comparable or better accuracy in simulation and decision-making than strong baselines.
• Significantly reduces decision reward degradation and improves worst-case accuracy under perturbations.
• Effectively mitigates tail risks, validated by improved Conditional Value at Risk (CVaR@5).