AI 용어집
인공지능 완전 사전
Conditional Value-at-Risk
A coherent risk measure that calculates the expected loss conditional on it exceeding a VaR threshold. It provides more complete information about distribution tails than VaR alone.
Distributional Dynamic Programming
An extension of dynamic programming that maintains and propagates value distributions rather than point estimates. It enables more robust planning by accounting for uncertainty in transitions and rewards.
Risk-Aware Policy
A decision strategy that optimizes a trade-off between expected return and risk exposure according to specified preferences. It adapts its behavior to avoid actions with overly dispersed return distributions.
Return Distribution
The complete probability distribution of future cumulative returns from a given state for a specific policy. Its characterization allows for a finer assessment of risks and uncertainties than simple expected value.
Risk-Adjusted Return
A performance measure that normalizes returns based on the risk incurred to allow for fair comparisons. It combines return expectation with a penalty or bonus based on their variability or skewness.
Distributional Actor-Critic
An actor-critic architecture where the critic estimates the complete value distribution rather than a simple scalar estimate. The actor then directly optimizes objectives based on this enriched distribution.
Coherent Risk Measures
A class of risk measures satisfying mathematical properties that ensure rational risk management. They include convexity, monotonicity, translation invariance, and positive homogeneity.