Definition
Reinforcement learning methods that use deep networks to represent values, policies, or models of the environment.
DQN combined convolutional networks with action-value learning to learn control policies from game screens. The 2015 paper is a representative example, not the origin of all deep reinforcement learning.
Deep representations expanded the range of inputs that could be handled, but increased training instability and resource requirements. Experimental performance in games cannot directly establish reliability in real-world robotics or other tasks.
Principles and impact
Problems solved
- Enabled reinforcement learning to handle high-dimensional perceptual inputs and complex function approximation.
Principles
- Deep networks approximate values or policies, while rewards from the environment contribute to learning updates.
Improvements
Not recorded
Significance
- Demonstrated the ability to combine perceptual representations with sequential decision-making.
History and applications
People
Volodymyr Mnih
Organizations
Not recorded
Events
Not recorded
Applications
Decision-making in games、Research on high-dimensional control
Limitations
- Sample requirements, training stability, and transfer between environments remain constraints.
Upstream and downstream relations
Upstream
Sources: Human-level Control through Deep Reinforcement Learning
Sources: Human-level Control through Deep Reinforcement Learning
Sources: Human-level Control through Deep Reinforcement Learning
Downstream
No direct downstream relations recorded