Global Convergence of Policy Gradient Methods to (Almost) Locally Optimal Policies.
Kaiqing ZhangAlec KoppelHao ZhuTamer BasarPublished in: SIAM J. Control. Optim. (2020)
Keyphrases
- global convergence
- optimal policy
- policy gradient methods
- convergence rate
- global optimum
- optimization methods
- markov decision processes
- reinforcement learning
- average reward
- convergence analysis
- convergence speed
- natural actor critic
- policy gradient
- decision problems
- dynamic programming
- state space
- long run
- actor critic
- partially observable markov decision processes
- policy iteration
- infinite horizon
- markov decision process
- finite state
- sufficient conditions
- hybrid algorithm
- optimization problems
- average cost
- initial state
- optimization method
- markov decision problems
- objective function
- particle swarm
- step size
- function approximation
- differential evolution
- simulated annealing