Search

The target net method has been the foundation of deep reinforcement learning since Deepmind first proposed it in 2015. Almost all the current popular reinforcement learning algorithms include target net. However, while the slowly updated target network improves the stability of the algorithm, it also reduces the performance of the algorithm. In this paper, the authors design a novel triple-network algorithm(TPN). TPN combines the temporal-difference(TD) algorithm and policy gradient(PG) theorem. Using three networks to estimate the state value(v), action value (q) , and policy(π). These networks have no primary or secondary distinction but are trained synchronously and influence each other. The author found that through this TPN architecture, the convergence and stability of the algorithm can be greatly improved without increasing the amount of calculation. Although it is only a basic framework at present. The calculation process of TPN is simple and easy to implement. Experiments prove that the convergence speed and stability of TPN in discrete cases are better than PPO. • Proposed a new deep reinforcement learning algorithm that does not rely on target networks. • Presented the complete derivation process of the TPN algorithm. • Verified the effectiveness of proposed algorithm through experiments. [ABSTRACT FROM AUTHOR]

The Wasserstein distance in the optimal transport (OT) model is being widely used in machine learning applications due to its mathematical and geometric properties. Since the entropy regularized OT is a convex problem, and for its large-scale problems, many algorithms have been proposed. Existing methods mainly adopt iterative random projection techniques to approximate the Wasserstein distance in the OT model. In this paper, we employ the entropic domain decomposition algorithm with mini-batch learning to calculate the entropy regularized OT problem and investigate the stability of this algorithm. The algorithm performs on different cells of the domain instead of using the projection technique and it is based on a mini-batch loss function. The corresponding solution in each iteration is derived from the dual problem of entropy regularized OT on each decomposed cell. Once the solutions to the dual problem on all decomposed cells are obtained, the corresponding transport plan in each iteration can be obtained. We show that the transport plan computed by its dual variables is admissible, and we also give the sub-gradient approximation result. In addition, we exploit the statistical concentration result of the subsampled mini-batch Wasserstein distance on decomposed cells. Finally, the numerical experiments validate that the proposed algorithm can save computational time further. [ABSTRACT FROM AUTHOR]

Deep Q-Networks algorithm (DQN) was the first reinforcement learning algorithm using deep neural network to successfully surpass human level performance in a number of Atari learning environments. However, divergent and unstable behaviour have been long standing issues in DQNs. The unstable behaviour is often characterised by overestimation in the Q -values, commonly referred to as the overestimation bias. To address the overestimation bias and the divergent behaviour, a number of heuristic extensions have been proposed. Notably, multi-step updates have been shown to drastically reduce unstable behaviour while improving agent's training performance. However, agents are often highly sensitive to the selection of the multi-step update horizon (n), and our empirical experiments show that a poorly chosen static value for n can in many cases lead to worse performance than single-step DQN. Inspired by the success of n -step DQN and the effects that multi-step updates have on overestimation bias, this paper proposes a new algorithm that we call 'Elastic Step DQN' (ES-DQN) to alleviate overestimation bias in DQNs. ES-DQN dynamically varies the step size horizon in multi-step updates based on the similarity between states visited. Our empirical evaluation shows that ES-DQN out-performs n -step with fixed n updates, Double DQN and Average DQN in several OpenAI Gym environments while at the same time alleviating the overestimation bias. [ABSTRACT FROM AUTHOR]