Jilles S. Dibangoye: Planning and Learning in Partially Observable Stochastic Games

Planning and Learning in Partially Observable Stochastic Games 

Today, real-life applications widely use Multi-Agent Systems (MAS), that is, groups of autonomous, interacting agents sharing a common environment, which they perceive through sensors and upon which they act with actuators, e.g., robots, drones, cars, power generators, smartphones, televisions, vacuum cleaners, washing machines. The increasing penetration of MASs in society will require a paradigm shift - from single-agent to multi-agent planning and Reinforcement Learning (RL) algorithms - leveraging on recent breakthroughs.

The main idea presented in this talk is that it is possible to reduce a partially observable stochastic game (POSG) to a fully observable stochastic game (SG). The resulting problem is then solved using generic algorithms based on recent advances in Artificial Intelligence. We shall discuss the system designer approach which eases these reformulations and provides a framework for establishing for the first time structural results for POSGs. We shall also present generic methods with theoretical guarantees for solving pure cooperative or competitive POSGs.

August 7, 2019 (Wednesday)
Room 205, Building E, Karlovo nám. 13


Jilles S. Dibangoye

Jilles Steeve Dibangoye is an associate professor at INSA Lyon, researcher in Artificial Intelligence at CITI lab, where he focuses on building machines mimicking and eventually exceeding human intelligence in INRIA team Chroma. See the list of publications to get to know his work.