Niccolò Turcato - Winning the AI Olympics Challenge With Model-Based Reinforcement Learning

In this talk, we present a Reinforcement Learning (RL) approach that we implemented to tackle the "AI Olympics With RealAIGym" competition held at IJCAI 2023. Our algorithm, named Monte-Carlo Probabilistic Inference for Learning COntrol (MC-PILCO) [Amadio et al., 2022], is a Model-Based (MB) RL algorithm that proved remarkably data-efficient in several low-dimensional benchmarks, such as a cart-pole, a ball & plate, and a Furuta pendulum, both in simulation and real setups. MC-PILCO exploits data collected by interacting with the system to derive a model of the system dynamics, and optimizes the policy by simulating the system, rather than optimizing the policy directly on the actual system. When considering physical systems, this kind of approach can be highly performing and more data-efficient than Model-Free (MF) solutions.

Niccolò Turcato

"AI Olympics with RealAIGym" Competition at IJCAI 2023