Catastrophic-risk-aware reinforcement learning with extreme-value-theory-based policy gradients.
Parisa DavarFrédéric GodinJose GarridoPublished in: CoRR (2024)
Keyphrases
- extreme value theory
- reinforcement learning
- optimal policy
- policy search
- markov decision process
- action selection
- partially observable environments
- actor critic
- markov decision processes
- policy gradient
- markov decision problems
- control policies
- partially observable
- action space
- function approximators
- state and action spaces
- state space
- risk assessment
- policy evaluation
- function approximation
- reinforcement learning problems
- reward function
- state action
- model free
- approximate dynamic programming
- decision problems
- policy iteration
- decision making
- rl algorithms
- control policy
- policy gradient methods
- risk management
- reinforcement learning algorithms
- partially observable markov decision processes
- dynamic programming
- risk measures
- temporal difference learning
- expected utility
- temporal difference
- average reward
- learning process
- opportunity cost
- agent learns
- risk averse
- infinite horizon
- gradient information
- control problems
- transition model
- partially observable domains
- machine learning
- agent receives
- model free reinforcement learning
- least squares
- supervised learning
- inverse reinforcement learning
- markov chain
- partially observable markov decision process
- learning problems
- finite state