Abstract:Reinforcement Learning,as a subject of study for over more than fifty years,investigates how an autonomous agent can learn what to do to maximize a numerical reward signal from interaction with the world by balancing exploration of the environment with exploitation of knowledge gained via evaluative feedback,without relying on exemplary supervision of an omniscient teacher or complete models of the environment.Deep learning is a cutting-edge approach to machine learning that concerns with using multi-layer artificial neural networks to learn the complicated representations that are expressed in terms of simpler ones.Currently,Deep Reinforcement Learning formed by combining modern reinforcement learning with deep learning is becoming a new research hotspot in the Artificial Intelligence community,and has made substantial breakthroughs in a variety of tasks-such as robot control,text recognition and games-requiring both rich perception of high-dimensional raw inputs and policy control.In particular,a state-of-the-art deep reinforcement learning model,termed Deep Q-Network,is able to perform human-level control using the same network architecture and hyper-parameters for handling problems approaching real-world complexity such as some Atari 2600 games.However,Deep Q-Network''s performance falls far below human level in situations that exist delayed rewards and require planning under uncertainty within long-time horizon to optimize strategies.This implies that Deep Q-Network is not good at controlling agents in strategic deep reinforcement learning tasks.To alleviate the issue,this paper proposes a novel deep reinforcement learning model by improving Deep Q-Network with recurrent neural networks based on visual attention mechanism.Two key ideas are included in the new model: (1) it uses recurrent neural networks consisting of two-layer gated recurrent units in order to remember more historical information of multiple time steps.This can make an agent exploit delayed feedback in time to guide its next action selection online.By using recurrent neural networks,the scale of input state is reduced from four stacked images to one current raw image.This can substantially reduce the state space;(2) the visual attention mechanism is used to adaptively focus attention on smaller but more valuable regions of an input image,and make agents control the process of learning near optimal policies more effectively.As a result,the number of parameters to be learned by a stochastic gradient descent method during training can be decreased sharply by introducing the visual attention mechanism.This can speed up the process of learning near optimal policies.This new model is actually equivalent to an encoder-decoder architecture,where the convolutional neural networks play an encoder role for extracting useful features,and the recurrent neural networks based on visual attention mechanism play the other.We used five challenging strategic tasks from the set of classic Atari 2600 games,i.e.,Seaquest,Alien,Gopher,Asteroids,and Gravitar,to verify the effectiveness of the new model.Experimental results show that artificial agents generated through our new model surpass DQN and its variant''s performance in terms of the average reward per episode,training speed,and policy stability on them,especially on the Seaquest and Gopher games.

A Model-Based Exploration Policy in Deep Q-Network

DQN with model-based exploration: efficient learning on environments with sparse rewards

State of the Art Control of Atari Games Using Shallow Reinforcement Learning

Sampling Efficient Deep Reinforcement Learning through Preference-Guided Stochastic Exploration

Random curiosity-driven exploration in deep reinforcement learning

A Novel Policy Based on Action Confidence Limit to Improve Exploration Efficiency in Reinforcement Learning

On the Importance of Exploration for Real Life Learned Algorithms

Careful at Estimation and Bold at Exploration

On the Convergence and Sample Complexity Analysis of Deep Q-Networks with $ε$-Greedy Exploration

Deep Q-learning Sampling Based on Advantages

Curiosity-Driven Variational Autoencoder for Deep Q Network

Improving Exploration in Evolution Strategies for Deep Reinforcement Learning via a Population of Novelty-Seeking Agents

Efficient and Scalable Exploration Via Estimation-Error

A Scalable Derivative-free Exploration Approach for Reinforcement Learning

A Comparative Study of Deep Reinforcement Learning Models: DQN vs PPO vs A2C

Distributional Reinforcement Learning for Efficient Exploration

A Deep Recurrent Q-Network Based on Visual Attention Mechanism

Policy Optimization with Model-based Explorations

A Human Mixed Strategy Approach to Deep Reinforcement Learning

Solving Deep Reinforcement Learning Tasks with Evolution Strategies and Linear Policy Networks

Langevin DQN