Using Reward Machines for High-Level Task Specification and Decomposition in Reinforcement Learning.
Rodrigo Toro IcarteToryn Q. KlassenRichard Anthony ValenzanoSheila A. McIlraithPublished in: ICML (2018)
Keyphrases
- reinforcement learning
- high level
- low level
- state space
- function approximation
- reward function
- eligibility traces
- decomposition method
- temporal difference
- formal specification
- reinforcement learning algorithms
- higher level
- conceptual model
- optimal control
- low level features
- markov decision processes
- lower level
- average reward
- total reward
- learning algorithm
- programming language
- supervised learning
- dynamic programming
- multi agent
- machine learning
- optimal policy
- sufficient conditions
- learning classifier systems
- source code
- mobile robot
- decomposition methods
- temporal abstractions
- partially observable environments