Reinforcement Learning for Finite Horizons (ReLeaF)

点击翻译按钮获取中文摘要

Reinforcement learning (RL) is a technique for learning how to take actions in an initially unknown environment in order to optimise an expected outcome, which is modelled through the notion of maximising an accumulative reward. Learning algorithms with goals written as temporal specifications have three key ingredients: the translation from the specification to appropriate finite automata; the translation of these finite automata to reward structures, such that a strategy that provides optimal rewards is guaranteed to provide optimal control; and a wrapper into a discounting scheme that, for appropriate parameters, will ensure that a learner converge to an optimal strategy.We will consider the RL problems for a popular specification language used in automation and motion planning, the finite horizon linear time temporal logic LTLf. In particular, we will study model-free RL algorithms, which are more suitable to real-world applications where the behaviour of the environment is hard to predict, than its model-based counterpart. We will propose learning algorithms that provide translations from finite horizon LTL to reward structures with formal guarantees of satisfying the given goals for environments modelled as Markov Decision Processes (MDPs). We will extend our techniques to infinite-state MDPs, including variations where formal guarantees can be provided -- like countable, finitely branching MDPs -- and study conditions for our techniques to provide guarantees in more general classes, such as smoothness guarantees for compact MDPs. We will complement these lines of research by looking at goals with constraints. This is effectively considering prioritised goals, where meeting safety constraints takes precedence, while other properties -- such as efficiency -- are considered as tie-breakers among strategies that provide the same safety guarantees.