Natural Policy Gradient Research Articles

Exploring Parameter Space in Reinforcement Learning

Abstract This paper discusses parameter-based exploration methods for reinforcement learning. Parameter-based methods perturb parameters of a general function approximator directly, rather than adding noise to the resulting actions. Parameter-based exploration unifies reinforcement learning and black-box optimization, and has several advantages over action perturbation. We review two recent parameter-exploring algorithms: Natural Evolution Strategies and Policy Gradients with Parameter-Based Exploration. Both outperform state-of-the-art algorithms in several complex high-dimensional tasks commonly found in robot control. Furthermore, we describe how a novel exploration method, State-Dependent Exploration, can modify existing algorithms to mimic exploration in parameter space.

Natural actor-critic with baseline adjustment for variance reduction

In this study, we discuss a baseline function for the estimation of a natural policy gradient with respect to variance, and demonstrate a condition in which an optimal baseline function that reduces the variance is equivalent to the state value function. However, outside of this condition, the state value could be considerably different from the optimal baseline. For such cases, an extended version of the NTD algorithm is proposed, where an auxiliary function is estimated to adjust the baseline, being state value estimates in the original NTD version, to the optimal baseline. The proposed algorithm is applied to simple MDPs and a challenging pendulum swing-up problem.

Natural Actor-Critic

In this paper, we suggest a novel reinforcement learning architecture, the Natural Actor-Critic. The actor updates are achieved using stochastic policy gradients employing Amari's natural gradient approach, while the critic obtains both the natural policy gradient and additional parameters of a value function simultaneously by linear regression. We show that actor improvements with natural policy gradients are particularly appealing as these are independent of coordinate frame of the chosen policy representation, and can be estimated more efficiently than regular policy gradients. The critic makes use of a special basis function parameterization motivated by the policy-gradient compatible function approximation. We show that several well-known reinforcement learning methods such as the original Actor-Critic and Bradtke's Linear Quadratic Q-Learning are in fact Natural Actor-Critic algorithms. Empirical evaluations illustrate the effectiveness of our techniques in comparison to previous methods, and also demonstrate their applicability for learning control on an anthropomorphic robot arm.

Off-Policy Natural Policy Gradient Method for a Biped Walking Using a CPG Controller

Referring to the mechanism of animals’ rhythmic movements, motor control schemes using a central pattern generator (CPG) controller have been studied. We previously proposed reinforcement learning (RL) called the CPG-actor-critic model, as an autonomous learning framework for a CPG controller. Here, we propose an off-policy natural policy gradient RL algorithm for the CPG-actor-critic model, to solve the “exploration-exploitation” problem by meta-controlling “behavior policy.” We apply this RL algorithm to an automatic control problem using a biped robot simulator. Computer simulation demonstrated that the CPG controller enables the biped robot to walk stably and efficiently based on our new algorithm.