Decision Awareness in Reinforcement Learning - End-to-End Optimization Approaches

Explore decision awareness in reinforcement learning through this 54-minute seminar presented by Pierre-Luc Bacon from Université de Montréal. Delve into the end-to-end perspective of optimizing learning systems for optimal decision-making, focusing on recent advances in model-based reinforcement learning. Examine control-oriented transition models using implicit differentiation and the application of neural ordinary differential equations for nonlinear trajectory optimization. Investigate computational challenges and scaling solutions, including efficient Jacobian factorization in forward mode automatic differentiation and novel constrained optimizers inspired by adversarial learning. Cover topics such as optimal model design, smooth Bellman optimality equations, bilevel optimization, continuous-time meta-learning, and gradient flow-based techniques.