Legacy neural critics provide no mechanism to weight value estimates by domain properties practitioners already know matter, where every distinction (e.g., trend direction), is learned from scratch and might not be guaranteed. Our work explores having a designable critic instead of a neural critic for deep reinforcement learning in systems: a temporal-difference-learned graph whose values are indexed by First-Order Logic-defined distinctions, turning each expressed domain property into an explicit inductive bias. For instance, rather than hoping the critic learns that Feature A trending upwards is an important value region, we declare it and the graph’s updates make it so.

About Abhishek Duttagupta

Abhishek Duttagupta is a PhD student at UC3M and IMDEA Networks, where he is part of the Network Data Science group and Resilient AI Networks group. His research interests are reinforcement learning, learning for networking systems.

Este evento se impartirá en inglés