Policy learning for time-bounded reachability in Continuous-Time Markov Decision Processes via doubly-stochastic gradient ascent.
Ezio BartocciLuca BortolussiTomás BrázdilDimitrios MiliosGuido SanguinettiPublished in: CoRR (2016)
Keyphrases
- markov decision processes
- state space
- optimal policy
- reinforcement learning
- partially observable
- gradient ascent
- average reward
- policy gradient
- policy iteration
- markov decision process
- finite state
- reinforcement learning algorithms
- partially observable markov decision processes
- learning algorithm
- stochastic games
- average cost
- action selection
- decision problems
- supervised learning
- action space
- function approximators
- computational complexity
- np hard