Project

Probabilistic Deep Reinforcement Learning for multi-agent systems

In this project, we are developing a probabilistic deep reinforcement learning framework that allows for predictive control of complex, dynamical systems. In particular, we aim at methods that can synthesize controllers in fast-paced, multi-agent settings that involve multiple robotic cars. To the date, deep reinforcement learning is mostly applied to single agent scenarios, e.g., a single robot grasping an object. However, in many critical application domains, real-time decisions will have to be made based on information about team-mates or adversaries. Another limitation of current deep reinforcement learning frameworks is that they generate only single “point estimates” as predictions and do not accurately model uncertainty. We will leverage recent theoretical insights into Bayesian Neural Networks to generate probabilistic predictions of the control system. Such a probabilistic model will allow us to generate “aggressive” driving maneuvers in which a robotic car goes to the limit of what it can achieve. In particular, we will use these insights to learn model-predictive control policies that can predict the internal state of the controlled car in the future, including the expected sensor variables and forces.

Funded by Intel AI Labs