Risk-aware Reachability and SSP

We present a novel approach to risk in reachability and stochastic shortest path (SSP) problems on Markov chains and decision processes. In particular, we consider the time to arrive at the goal states and the total accumulated cost on the way, respectively, to quantify risk. Motivated by our previous work, we employ conditional value-at-risk (CVaR) as measure of risk. In essence, CVaR quantifies risk as the expectation of the worst p-quantile. We present ongoing work on the computational complexity of the respective decision problems, structure of optimal strategies, and, in contrast to our previous work, also a value iteration algorithm.

Authors

• Tobias Meggendorfer

Keywords

No keywords are indexed for this paper.