Robust Risk Under Evolving Uncertainty: A Wasserste... | AI Research

Robust Risk Under Evolving Uncertainty: A Wasserstein Counterpart of the Entropic Value-at-Risk addresses the challenge of how an agent should balance caution and boldness while learning about its environment. The authors, Deep Kumar Ganguly and Jan Křetínský, propose a new risk measure that adjusts its level of conservatism based on the agent's current belief, ensuring safety even when facing potential catastrophes that a standard model might deem impossible.

The Problem with Current Risk Measures

Existing approaches to risk-sensitive control often rely on relative-entropy (KL) balls to define "ambiguity sets"—the range of alternative models an agent considers to protect itself. However, the authors note that these balls have a "blind spot": they cannot account for catastrophic outcomes that the nominal model assigns zero probability. If an agent assumes a disaster is impossible, a relative-entropy-based controller will ignore that risk entirely, even as it learns. This leaves the agent vulnerable to rare but fatal events.

The Wasserstein Approach

To fix this, the authors introduce the Wasserstein entropic value-at-risk (WEVaR). By using an optimal-transport (Wasserstein) ball instead of a relative-entropy ball, the measure accounts for the physical "distance" to a catastrophic state. Because this geometry measures how far mass must move to reach a different outcome, it allows the agent to hedge against reachable disasters, even if the nominal model initially assigns them zero probability. The authors prove that this measure is coherent and fits into the existing risk hierarchy, providing a more robust safety net than previous methods.

Evolving Caution via Belief Entropy

The paper proposes a dynamic-programming operator where the ambiguity radius is driven by the entropy of the agent's belief. When the agent is ignorant (high entropy), the radius is large, forcing the agent to be cautious. As the agent gathers data and its belief sharpens (low entropy), the radius contracts, allowing the agent to become bolder. This creates a "safety switch" where the agent automatically transitions from a conservative strategy to a more efficient one as its confidence increases. The authors demonstrate that this approach provides a "safety sandwich," keeping the agent's value between a maximally cautious floor and an ideal, type-aware oracle.

Numerical Verification and Performance

The authors verified their theoretical findings using Gurobi to solve the robust-optimization problems. Their experiments on a five-state corridor confirm that the Wasserstein-based approach successfully accounts for catastrophes that the entropic measure ignores. The results show that the price of robustness—the cost of being cautious—is temporary and naturally dissipates as the agent identifies the environment. The authors note that this closed-form operator reduces the computational cost per state and belief to $O(|\mathcal{S}|^2)$, making it a practical tool for sequential decision-making under uncertainty.