# Deep Q-Network

> artificial intelligence algorithm made by Google_Deepmind in London

**Wikidata**: [Q20722857](https://www.wikidata.org/wiki/Q20722857)  
**Source**: https://4ort.xyz/entity/deep-q-network

## Summary
Deep Q-Network (DQN) is an artificial intelligence algorithm developed by Google DeepMind in London that combines deep learning with reinforcement learning to enable agents to learn optimal policies from high-dimensional sensory inputs.

## Key Facts
- Developed by Google DeepMind in London
- Combines deep neural networks with Q-learning
- First algorithm to successfully integrate deep learning with reinforcement learning
- Published in the seminal 2015 Nature paper "Human-level control through deep reinforcement learning"
- Achieved human-level performance on many Atari 2600 games using raw pixel inputs
- Uses experience replay to break correlation between consecutive samples
- Employs a target network to stabilize training
- Instance of artificial neural network, software library, and algorithm
- Subclass of artificial neural network

## FAQs
### Q: What is Deep Q-Network used for?
A: Deep Q-Network is used for reinforcement learning tasks where an agent must learn optimal behavior through trial and error, particularly in environments with high-dimensional state spaces like video games, robotics, and autonomous systems.

### Q: How does Deep Q-Network differ from traditional Q-learning?
A: Unlike traditional Q-learning which uses a table to store state-action values, DQN uses a deep neural network to approximate the Q-function, enabling it to handle continuous and high-dimensional state spaces that would be impossible with tabular methods.

### Q: What makes Deep Q-Network significant in AI research?
A: DQN was groundbreaking because it was the first algorithm to demonstrate that deep learning could be successfully combined with reinforcement learning to achieve human-level performance on complex tasks directly from raw sensory inputs.

## Why It Matters
Deep Q-Network represents a pivotal breakthrough in artificial intelligence that bridged the gap between deep learning and reinforcement learning. Before DQN, reinforcement learning algorithms struggled with high-dimensional inputs like images, while deep learning excelled at processing such data but lacked the ability to learn through interaction and reward. DQN solved this by using convolutional neural networks to process raw pixel inputs and Q-learning to make decisions, enabling agents to learn directly from sensory data without manual feature engineering. This innovation opened new possibilities for AI systems that can learn complex behaviors from scratch in real-world environments, influencing applications from game playing to robotics and autonomous vehicles. The algorithm's success on Atari games demonstrated that AI could achieve human-level performance on challenging tasks through end-to-end learning, establishing a foundation for subsequent advances in deep reinforcement learning.

## Notable For
- First algorithm to achieve human-level performance on Atari 2600 games using raw pixel inputs
- Pioneered the use of experience replay in deep reinforcement learning
- Introduced the target network technique to stabilize training
- Published in Nature, one of the most prestigious scientific journals
- Sparked the deep reinforcement learning revolution in AI research
- Demonstrated end-to-end learning from pixels to actions without manual feature engineering

## Body
Deep Q-Network was developed by Google DeepMind researchers in London, with the seminal paper published in Nature in February 2015. The algorithm combines Q-learning with deep neural networks, specifically using convolutional layers to process visual inputs. DQN processes raw pixel data from Atari 2600 games at 60 frames per second, using the same network architecture and hyperparameters across all 49 games tested. The algorithm employs experience replay, storing agent experiences (state, action, reward, next state) in a replay buffer and sampling mini-batches randomly to break correlation between consecutive samples. A target network, which is a periodically updated copy of the main Q-network, provides stable Q-value targets during training. DQN achieved performance comparable to or exceeding human experts on 29 of the 49 Atari games tested, with some games showing more than double the human score. The algorithm uses an ε-greedy exploration strategy, starting with ε=1.0 and linearly decaying to ε=0.1 over the first million frames. Training typically requires 200 million frames (about 38 days of game experience) to reach peak performance. The success of DQN established deep reinforcement learning as a viable approach for complex decision-making tasks and influenced subsequent algorithms like Double DQN, Dueling DQN, and Rainbow.

## Schema Markup
```json
{
  "@context": "https://schema.org",
  "@type": "Thing",
  "name": "Deep Q-Network",
  "description": "Artificial intelligence algorithm combining deep learning with reinforcement learning, developed by Google DeepMind in London",
  "url": "https://deepmind.com/research/open-source/dqn",
  "sameAs": [
    "https://www.wikidata.org/wiki/Q21593519",
    "https://en.wikipedia.org/wiki/Deep_Q-learning"
  ],
  "additionalType": "artificial neural network"
}

## References

1. [Source](https://github.com/deepmind/dqn)
2. [Source](https://misovalko.github.io/)