| Definition | The Procgen Benchmark is a collection of 16 dynamic, procedurally generated environments that push the boundries of how quickly reinforcement learning agents can acquire generalizable skills in randomized scenarios. |
| No of environments | 16 |
The Procgen Benchmark presents a suite of 16 procedurally generated environments, such as CoinRun and BossFight, designed to evaluate how swiftly reinforcement learning agents can acquire generalizable skills. Operating at thousands of steps per second on a single core, these environments allow researchers to effectively measure the learning efficiency and adaptability of various algorithms. This benchmark is particularly valuable for experimentation in dynamic settings where agents must navigate randomized scenarios rather than depend on memorizing fixed sequences. For instance, in CoinRun, players must avoid obstacles and enemies to reach a goal, while in BossFight, they need to dodge attacks from a larger adversary, showcasing different skill sets.
In contrast to static environments, these procedurally generated spaces introduce variability in design and challenges, ensuring each run is unique. This inherent unpredictability enhances the robustness of the learning process and fosters the development of agents that can adapt to unforeseen circumstances. For example, in CaveFlyer, players navigate a network of caves to reach an exit, facing different layouts with each attempt, while in FruitBot, agents must discern between collectible fruits and non-fruit objects. This flexibility allows researchers to easily customize the environments, modifying the underlying code to create new challenges, with the codebase for each environment remaining concise and manageable.
Supporting the environments is a comprehensive ecosystem of resources, including detailed documentation and example scripts that help users leverage the benchmark effectively. The Procgen Benchmark has become an essential tool within the reinforcement learning community, facilitating competitions that focus on sample efficiency and generalization. By emphasizing dynamic learning through engaging scenarios, such as Leaper and Plunder, this benchmark not only aids in evaluating existing algorithms but also expands the frontiers of what can be achieved in training RL agents across a variety of contexts.
| Example observations | |
| Example actions | |
| Example reward |