scholarly journals Off-Policy Deep Reinforcement Learning by Bootstrapping the Covariate Shift

2019 ◽  
Vol 33 ◽  
pp. 3647-3655
Author(s):  
Carles Gelada ◽  
Marc G. Bellemare
Keyword(s):  
Reinforcement Learning ◽  
Learning Algorithm ◽  
Theoretical Perspective ◽  
Empirical Evaluation ◽  
Nonlinear Function ◽  
Policy Learning ◽  
Extensive Evaluation ◽  
Probability Simplex ◽  
Performance Gains ◽  
Nonlinear Function Approximation

In this paper we revisit the method of off-policy corrections for reinforcement learning (COP-TD) pioneered by Hallak et al. (2017). Under this method, online updates to the value function are reweighted to avoid divergence issues typical of off-policy learning. While Hallak et al.’s solution is appealing, it cannot easily be transferred to nonlinear function approximation. First, it requires a projection step onto the probability simplex; second, even though the operator describing the expected behavior of the off-policy learning algorithm is convergent, it is not known to be a contraction mapping, and hence, may be more unstable in practice. We address these two issues by introducing a discount factor into COP-TD. We analyze the behavior of discounted COP-TD and find it better behaved from a theoretical perspective. We also propose an alternative soft normalization penalty that can be minimized online and obviates the need for an explicit projection step. We complement our analysis with an empirical evaluation of the two techniques in an off-policy setting on the game Pong from the Atari domain where we find discounted COP-TD to be better behaved in practice than the soft normalization penalty. Finally, we perform a more extensive evaluation of discounted COP-TD in 5 games of the Atari domain, where we find performance gains for our approach.

scholarly journals Metatrace Actor-Critic: Online Step-Size Tuning by Meta-gradient Descent for Reinforcement Learning Control

2019 ◽  
Author(s):  
Kenny Young ◽  
Baoxiang Wang ◽  
Matthew E. Taylor
Keyword(s):  
Reinforcement Learning ◽  
Gradient Descent ◽  
Single Agent ◽  
Nonlinear Function ◽  
Step Size ◽  
Experience Replay ◽  
Training Examples ◽  
The Cost ◽  
Single Agent Learning ◽  
Nonlinear Function Approximation

Reinforcement learning (RL) has had many successes, but significant hyperparameter tuning is commonly required to achieve good performance. Furthermore, when nonlinear function approximation is used, non-stationarity in the state representation can lead to learning instability. A variety of techniques exist to combat this --- most notably experience replay or the use of parallel actors. These techniques stabilize learning by making the RL problem more similar to the supervised setting. However, they come at the cost of moving away from the RL problem as it is typically formulated, that is, a single agent learning online without maintaining a large database of training examples. To address these issues, we propose Metatrace, a meta-gradient descent based algorithm to tune the step-size online. Metatrace leverages the structure of eligibility traces, and works for both tuning a scalar step-size and a respective step-size for each parameter. We empirically evaluate Metatrace for actor-critic on the Arcade Learning Environment. Results show Metatrace can speed up learning, and improve performance in non-stationary settings.

scholarly journals Computational Design of Modular Robots Based on Genetic Algorithm and Reinforcement Learning

Symmetry ◽  
2021 ◽  
Vol 13 (3) ◽  
pp. 471
Author(s):  
Jai Hoon Park ◽  
Kang Hoon Lee
Keyword(s):  
Genetic Algorithm ◽  
Reinforcement Learning ◽  
Design Space ◽  
Learning Algorithm ◽  
Computational Design ◽  
Computational Method ◽  
Learning Ability ◽  
Modular Robots ◽  
Control Mechanisms ◽  
Candidate Structure

Designing novel robots that can cope with a specific task is a challenging problem because of the enormous design space that involves both morphological structures and control mechanisms. To this end, we present a computational method for automating the design of modular robots. Our method employs a genetic algorithm to evolve robotic structures as an outer optimization, and it applies a reinforcement learning algorithm to each candidate structure to train its behavior and evaluate its potential learning ability as an inner optimization. The size of the design space is reduced significantly by evolving only the robotic structure and by performing behavioral optimization using a separate training algorithm compared to that when both the structure and behavior are evolved simultaneously. Mutual dependence between evolution and learning is achieved by regarding the mean cumulative rewards of a candidate structure in the reinforcement learning as its fitness in the genetic algorithm. Therefore, our method searches for prospective robotic structures that can potentially lead to near-optimal behaviors if trained sufficiently. We demonstrate the usefulness of our method through several effective design results that were automatically generated in the process of experimenting with actual modular robotics kit.

Sign in / Sign up

Export Citation Format

Share Document