Non-Deterministic Policy Improvement Stabilizes Approximated Reinforcement Learning.
Wendelin BöhmerRong GuoKlaus ObermayerPublished in: CoRR (2016)
Keyphrases
- reinforcement learning
- optimal policy
- policy search
- markov decision process
- state space
- action selection
- partially observable environments
- markov decision processes
- reinforcement learning problems
- partially observable
- control policies
- function approximation
- policy gradient
- reinforcement learning algorithms
- partially observable domains
- dynamic programming
- deterministic domains
- control policy
- function approximators
- learning algorithm
- policy iteration
- action space
- agent learns
- model free
- markov decision problems
- approximate dynamic programming
- significant improvement
- continuous state spaces
- reward function
- policy evaluation
- learning process
- average cost
- average reward
- state and action spaces
- control problems
- linear combination
- sufficient conditions