Arcade Learning Environment (ALE) / Atari 100k

The standard platform for testing reinforcement learning on Atari 2600 games; Atari 100k is its low-sample variant.

The Arcade Learning Environment (ALE) was proposed by Bellemare and colleagues, with the paper published in JAIR in 2013. Built on the Stella emulator, it wraps over a hundred Atari 2600 games behind a uniform interface: the agent sees screen pixels, chooses joystick actions, and receives the game score as reward. After DQN reached human-level performance on it in 2015, it became the most widely used benchmark in deep reinforcement learning, and is now maintained by the Farama Foundation. Atari 100k comes from the 2019 SimPLe paper: across 26 games, the agent is allowed only about 100,000 steps of interaction (roughly two hours of gameplay), specifically to test sample efficiency, and world-model methods such as DreamerV3 and DIAMOND are commonly compared on it. A great deal of the reinforcement-learning and world-model technology used in embodied AI today was first validated here.

ExampleUnder the Atari 100k setting, an agent playing Breakout gets only about two hours' worth of interaction data to learn from, and its score is then compared, normalized, against human performance.

Also called

Arcade Learning Environment, Atari 100k, Atari 2600 benchmark

Related

Reinforcement Learning · Deep Q-Network · Sample Efficiency · World Model · DreamerV3 · DIAMOND

Sources

The Arcade Learning Environment: An Evaluation Platform for General Agents (Bellemare et al.)
Arcade Learning Environment 文档（Farama Foundation） (Chinese)
Model-Based Reinforcement Learning for Atari (SimPLe, Kaiser et al., 2019)