Deep Reinforcement Learning: A Survey.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Reinforcement Learning

Background:

• Deep reinforcement learning (DRL) merges deep learning's feature representation with reinforcement learning's decision-making for end-to-end control.
• DRL has advanced significantly in tasks with high-dimensional inputs and optimal decision-making over the past decade.
• Challenges persist in DRL, particularly in sample-limited, sparse-reward, and multi-agent learning control tasks.

Purpose of the Study:

• To provide a comprehensive overview of the fundamental theories, key algorithms, and primary research domains of DRL.
• To summarize advances in value-based, policy-based, and maximum entropy-based DRL algorithms.
• To analyze and discuss future research directions in DRL.

Main Methods:

• Review of existing literature on Deep Reinforcement Learning.
• Categorization and summarization of DRL algorithms (value-based, policy-based, maximum entropy-based).
• Analysis of current challenges and future research trends in DRL.

Main Results:

• Significant advances in DRL have been observed across various complex tasks.
• Various solutions and theories have been proposed to address DRL's inherent challenges.
• Deep learning has spurred advancements in subfields like hierarchical and multi-agent reinforcement learning.

Conclusions:

• DRL offers powerful capabilities for learning control from high-dimensional data.
• Continued research is essential to overcome limitations in sample efficiency, reward sparsity, and multi-agent coordination.
• The field is poised for further development, driven by deep learning integration and exploration of new algorithmic approaches.