The Next Input updates The Next Input-updates

More on Dota 2 More on Dota 2

Our Dota 2 result shows that self-play can catapult the performance of machine learning systems from far below human level to superhuman, given sufficient compute. In the span of a month, our system went from barely matching a high-ranked player to beating the top pros and has continued to improve since then. Supervised deep learning systems can only be as good as their training datasets, but in self-play systems, the available data improves automatically as the agent gets better. Our Dota 2 result shows that self-play can catapult the performance of machine learning systems from far below human level to superhuman, given sufficient compute. In the span of a month, our system went from barely matching a high-ranked player to beating the top pros and has continued to improve since then. Supervised deep learning systems can only be as good as their training datasets, but in self-play systems, the available data improves automatically as the agent gets better.

Dota 2 Dota 2

Rewatch live event Rewatch live event

Gathering human feedback Gathering human feedback

View code(opens in a new window) View code(opens in a new window)

Better exploration with parameter noise Better exploration with parameter noise

Read code(opens in a new window)Read paper(opens in a new window) Read code(opens in a new window)Read paper(opens in a new window)

Proximal Policy Optimization Proximal Policy Optimization

View code(opens in a new window)Read paper(opens in a new window) View code(opens in a new window)Read paper(opens in a new window)

Robust adversarial inputs Robust adversarial inputs

We’ve created images that reliably fool neural network classifiers when viewed from varied scales and perspectives. This challenges a claim from last week that self-driving cars would be hard to trick maliciously since they capture images from multiple scales, angles, perspectives, and the like. We’ve created images that reliably fool neural network classifiers when viewed from varied scales and perspectives. This challenges a claim from last week that self-driving cars would be hard to trick maliciously since they capture images from multiple scales, angles, perspectives, and the like.

Hindsight Experience Replay Hindsight Experience Replay

Teacher–student curriculum learning Teacher–student curriculum learning

Read paper(opens in a new window) Read paper(opens in a new window)

Faster physics in Python Faster physics in Python

We’re open-sourcing a high-performance Python library for robotic simulation using the MuJoCo engine, developed over our past year of robotics research. We’re open-sourcing a high-performance Python library for robotic simulation using the MuJoCo engine, developed over our past year of robotics research.

Learning from human preferences Learning from human preferences

Read paper(opens in a new window) Read paper(opens in a new window)

Learning to cooperate, compete, and communicate Learning to cooperate, compete, and communicate

View code(opens in a new window)Read paper(opens in a new window) View code(opens in a new window)Read paper(opens in a new window)

UCB exploration via Q-ensembles UCB exploration via Q-ensembles