New Directions in RL: TD(lambda), aggregation, seminorm projections, free-form sampling (from 2014) скачать в хорошем качестве

New Directions in RL: TD(lambda), aggregation, seminorm projections, free-form sampling (from 2014)

This lecture explores three interrelated research directions in approximate dynamic programming and reinforcement learning: 1. Seminorm projections (unifying projected equation and aggregation approaches), generalized Bellman equations (multistep equations with state-dependent weights; the TD(lambda) equation is an example), and free form sampling (a flexible alternative to single long trajectory simulation) 2 Aggregation and seminorm projected equations 3 Simulation-based implementation of iterative and matrix inversion methods using free-form sampling. Part of this material has appeared in varying degrees of detail in my 2012 DP book (Vol. II), and my 2022 Abstract DP book. Slides at http://www.mit.edu/~dimitrib/Gen_Bell...