Multisource Transfer Double DQN Based on Actor Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Reinforcement Learning

Background:

• Deep Reinforcement Learning (RL) combines trial-and-error learning with deep learning's feature expression.
• Deep Q-Networks (DQN) face learning inefficiencies due to numerous parameters from constant environmental interaction.
• Action overestimation in DQN leads to error accumulation, impacting performance.

Purpose of the Study:

• To propose a novel deep reinforcement learning approach, Multisource Transfer Double Deep Q-Network (MTDDQN).
• To enhance learning efficiency and accuracy in RL agents by integrating transfer learning.
• To address action overestimation issues inherent in DQN.

Main Methods:

• Integration of transfer learning techniques (policy mimic, feature regression) with deep RL.
• Utilizing Double DQN to train the transfer network, mitigating action overestimation.
• Application of a multisource transfer learning mechanism to ensure task correlation and prevent negative transfer.

Main Results:

• MTDDQN demonstrated human-like actor learning transfer capabilities.
• The proposed method achieved improved learning efficiency compared to DQN and Double DQN.
• Experiments on the Atari2600 platform validated the feasibility and performance of MTDDQN.

Conclusions:

• MTDDQN effectively enhances learning efficiency and testing accuracy in deep reinforcement learning.
• The integration of multisource transfer learning and Double DQN addresses key limitations of existing methods.
• MTDDQN offers a promising approach for more capable and efficient RL agents.