CVaR

The researchers designed a novel decision rule that can map each credal set to a single predictive distribution, specifically for CVaR minimization, striking a balance between conservativeness and usability.

Conditional Value at Risk (CVaR) is defined as the expected value of the tail portion of the loss distribution exceeding a certain confidence level α, expressed mathematically as CVaR_α = E[L | L ≥ VaR_α], and possesses optimization properties such as convexity and subadditivity.

In risk-averse machine learning, besides CVaR, there also exist methods such as variance penalties, Wasserstein distance constraints and distributionally robust optimization (DRO).

A learner estimating the loss distribution from limited data inevitably has bias, causing the estimate of CVaR to deviate from the true value, thereby weakening or even reversing the actual effect of risk aversion.

CVaR possesses subadditivity and is a coherent risk measure.

The solution proposed in the paper applies CVaR to the immediate cost over the belief at each step, rather than modifying the value function, thereby preserving the standard expected cumulative return as the overall objective and maintaining the standard MDP structure.

Existing CVaR-based risk-averse methods apply static or dynamic CVaR to the value function to capture trajectory-level risk, but neglect the risk within the belief.

Introducing risk-sensitivity by modifying the value function destroys the standard structure of the original problem, forcing researchers to design entirely new custom algorithms instead of reusing existing expectation-based planners.

CVaR_α measures the average loss of a random variable in the worst (1-α) fraction of cases; as α→1 it degenerates to the expected value, and as α→0 it approaches the worst-case cost.

Commonly used metrics for assessing tail risk include Conditional Value at Risk (CVaR), worst-case analysis, and adversarial testing.