Computing monotone policies for Markov decision processes: a nearly-isotonic penalty approach.
Robert MattilaCristian R. RojasVikram KrishnamurthyBo WahlbergPublished in: CoRR (2017)
Keyphrases
- state and action spaces
- markov decision processes
- optimal policy
- markov decision problems
- discounted reward
- action space
- markov decision process
- decision processes
- state space
- average cost
- average reward
- finite state
- reinforcement learning
- reward function
- decentralized control
- dynamic programming
- total reward
- planning under uncertainty
- policy iteration
- partially observable markov decision processes
- stationary policies
- decision problems
- finite horizon
- reinforcement learning algorithms
- infinite horizon
- expected reward
- partially observable
- macro actions
- risk sensitive
- decision theoretic planning
- long run
- transition matrices
- factored mdps
- control policies
- multistage
- reachability analysis
- objective function
- policy iteration algorithm
- boolean functions
- learning algorithm
- interval estimation
- stochastic shortest path
- model based reinforcement learning
- markov chain
- sufficient conditions
- probabilistic planning
- initial state