We investigate effective methods to generate policies in stochastic environments, typically modeled as MDPs.
We investigate effective methods to learn interpretable models from data.
Everybody should be able to tell their machines what to do. Software synthesis is the problem of constructing software (i.e. a program) from user intent. Our work investigates automatic construction of reactive programs that run in perpetuity, and also programs that terminate. We use Linear Temporal Logic (LTL) as specification language.
In Fully Observable Non-deterministic (FOND) Planning, the effects of the actions that that an agent may perform are non-deterministic. We study Fully Observable Non-deterministic Planning with Temporally Extended Goals (AAAI-17 paper). Our techniques compute finite, and infinite plans whose execution satisfy a given property that is temporally extended, and that are robust to the non-determinism in the effect of the actions.