No More Hand-Tuning Rewards: Masked Constrained Policy Optimization for Safe Reinforcement Learning.
Stef Van HavermaetYara KhalufPieter SimoensPublished in: AAMAS (2021)
Keyphrases
- reinforcement learning
- optimal policy
- markov decision processes
- policy search
- reward function
- state space
- markov decision process
- concave convex procedure
- action selection
- control policy
- reinforcement learning algorithms
- function approximation
- policy iteration
- optimization algorithm
- control policies
- approximate dynamic programming
- total reward
- partially observable
- action space
- global optimization
- multi agent
- fine tuning
- function approximators
- partially observable markov decision processes
- temporal difference
- machine learning
- partially observable environments
- actor critic
- policy evaluation
- policy gradient
- model free
- state action
- markov decision problems
- decision problems
- finite state
- average reward
- optimal control
- continuous state spaces
- optimization problems
- reinforcement learning problems
- dynamic programming
- learning algorithm
- partially observable domains
- average cost
- discounted reward
- agent learns
- tuning parameters
- long run
- learning classifier systems
- constrained optimization