VESCAA: Verifikovatelná a efektivní syntéza kontrolerů

VESCAA: Verifiable and Efficient Synthesis of Agent Controllers

grant

Decision making under uncertainty; controller design; safety and scalalbility; inductive synthesis; reinforcement learning, risk-aware learning;

Many modern computing systems can be seen as (semi)-autonomous agents interacting with their environment. The agent's behaviour is determined by a controller that necessarily needs to deal with uncertainties including unpredictability of the environment and the imprecision of data gathered about its current state. There exists a multitude of approaches to automated controller design, however, they all tackle the safety-scalability gap: scalability limits the complexity of the problems that can be handled and safety ensures that agent operates in a safe and interpretable way. There are two principal approaches: formal methods prioritize safety and reinforcement learning prioritizes scalability. The project aims at developing theoretical foundation and synthesis algorithms that reduce this gap and thus improve their practical applicability. The key idea is to adapt, further develop and synergically integrate two emerging paradigms:  inductive synthesis improving the scalability of correct-by-construction design techniques and risk-aware learning improving the safety guarantees.

Češka Milan, doc. RNDr., Ph.D. (DITS) – research leader
Andriushchenko Roman, Ing. (DITS)