Human-level control through deep reinforcement learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Reinforcement learning (RL) optimizes agent control based on psychological and neuroscientific principles.
• Real-world RL requires agents to derive efficient representations from high-dimensional sensory inputs for generalization.
• Existing RL agents are limited to handcrafted features or low-dimensional, fully observed states.

Purpose of the Study:

• To develop a novel artificial agent capable of end-to-end reinforcement learning from high-dimensional sensory inputs.
• To overcome limitations of previous RL agents in complex, real-world scenarios.
• To bridge the gap between raw sensory data and effective decision-making in artificial agents.

Main Methods:

• Utilized advances in deep neural network training to create a deep Q-network agent.
• Employed end-to-end reinforcement learning, processing only raw pixels and game scores as inputs.
• Tested the agent on a diverse set of 49 classic Atari 2600 games.

Main Results:

• The deep Q-network agent surpassed all previous algorithms in Atari 2600 games.
• Achieved performance comparable to professional human games testers across the tested games.
• Demonstrated successful learning and generalization directly from high-dimensional visual input.

Conclusions:

• The deep Q-network represents a significant advancement in artificial intelligence, enabling learning from raw sensory data.
• This approach bridges the divide between high-dimensional inputs and actions, creating versatile agents.
• The agent's success across a variety of challenging tasks highlights the potential of deep reinforcement learning.