Shapley value-driven multi-modal deep reinforcement learning for complex decision-making.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Deep Reinforcement Learning (DRL) excels in sequential decision-making but struggles with complex, multimodal real-world environments.
• Traditional DRL faces limitations due to single-modal data, insufficient samples, and representation conflicts, hindering applications like autonomous driving.

Purpose of the Study:

• To introduce a novel Multi-Modal Deep Reinforcement Learning (MMDRL) framework integrating DRL with multimodal learning.
• To enhance information extraction, utilization, and decision-making in complex environments.
• To address challenges in multimodal data integration, sample scarcity, and policy optimization.

Main Methods:

• Developed an MMDRL framework combining deep reinforcement learning with multimodal learning.
• Implemented a knowledge-based sample augmentation technique to enrich training data and improve generalization.
• Conceptualized environmental perception as a multi-agent problem, using Shapley values for modality contribution evaluation and policy optimization.

Main Results:

• The proposed MMDRL framework effectively integrates multimodal information for enhanced decision-making.
• Knowledge-based sample augmentation significantly improved model generalization capabilities.
• Shapley value-based optimization reduced computational complexity and improved policy decisions in continuous action spaces.

Conclusions:

• MMDRL offers a robust solution for complex decision-making tasks in real-world multimodal scenarios.
• The framework demonstrates superior accuracy and efficiency in agent decision-making, validated on MuJoCo and Atari benchmarks.
• This research advances DRL applications, providing practical methods for autonomous systems and other intricate domains.