Metatrace: Online Step-size Tuning by Meta-gradient Descent for Reinforcement Learning Control.
Kenny YoungBaoxiang WangMatthew E. TaylorPublished in: CoRR (2018)
Keyphrases
- step size
- cost function
- reinforcement learning
- temporal difference
- stochastic gradient descent
- convergence rate
- convergence speed
- optimal control
- online learning
- steady state error
- control system
- evolutionary programming
- hessian matrix
- gradient method
- control policy
- optimal policy
- state space
- steepest descent method
- line search
- adaptive filter
- variable step size
- faster convergence
- action selection
- control method
- model free
- conjugate gradient
- function approximation
- loss function
- image compression
- quantization step
- multiresolution
- objective function
- image processing