Aiming at the contradiction between exploration and exploitation in deep reinforcement learning, this paper proposes “reward-based exploration strategy combined with Softmax action selection” (RBE-Softmax) as a dynamic exploration strategy to guide the agent to learn. The superiority of the proposed method is that the characteristic of agent's learning process is utilized to adapt exploration parameters online, and the agent is able to select potential optimal action more effectively. The proposed method is evaluated in discrete and continuous control tasks on OpenAI Gym, and the empirical evaluation results show that RBE-Softmax method leads to statistically-significant improvement in the performance of deep reinforcement learning algorithms.

The commonly used Deep Q Networks is known to overestimate action values under certain conditions. It's also proved that overestimations do harm to performance, which might cause instability and divergence of learning. In this paper, we present the Deep Sarsa and Q Networks (DSQN) algorithm, which can considered as an enhancement to the Deep Q Networks algorithm. First, DSQN algorithm takes advantage of the experience replay and target network techniques in Deep Q Networks to improve the stability of neural networks. Second, double estimator is utilized for Q-learning to reduce overestimations. Especially, we introduce Sarsa learning to Deep Q Networks for removing overestimations further. Finally, DSQN algorithm is evaluated on cart-pole balancing, mountain car and lunarlander control task from the OpenAI Gym. The empirical evaluation results show that the proposed method leads to reduced overestimations, more stable learning process and improved performance.