Bootstrapped model learning and error correction for planning with uncertainty in model-based RL

Having access to a forward model enables the use of planning algorithms such\nas Monte Carlo Tree Search and Rolling Horizon Evolution. Where a model is\nunavailable, a natural aim is to learn a model that reflects accurately the\ndynamics of the environment. In many situations it might not be possible and\nminimal glitches in the model may lead to poor performance and failure. This\npaper explores the problem of model misspecification through uncertainty-aware\nreinforcement learning agents. We propose a bootstrapped multi-headed neural\nnetwork that learns the distribution of future states and rewards. We\nexperiment with a number of schemes to extract the most likely predictions.\nMoreover, we also introduce a global error correction filter that applies\nhigh-level constraints guided by the context provided through the predictive\ndistribution. We illustrate our approach on Minipacman. The evaluation\ndemonstrates that when dealing with imperfect models, our methods exhibit\nincreased performance and stability, both in terms of model accuracy and in its\nuse within a planning algorithm.\n