Curiosity-Driven Testing for Sequential Decision-Making Process

Testing a sequential decision-making system — a reinforcement learning policy driving a car, say — is not like testing a function. A failure is a trajectory, the state space is enormous, and random search wastes almost all of its budget re-visiting states that behave the same way.

CureFuzz treats novelty as the thing worth optimizing: a curiosity mechanism steers the black-box fuzzer toward states unlike those already seen, and a multi-objective seed selection strategy balances that exploration against the likelihood of actually triggering a failure. Across several sequential decision-making systems it uncovers more distinct failure-triggering states than existing approaches, and does so without any access to the policy internals.