Reinforcement learning entails an artificial intelligence (AI) method that utilizes rewards or even punishments in driving agents towards the direction of specific objectives. In fact, the same technique was used in training the systems famous for defeating Alpha Go world champions as well as mastering Valve’s Dota2.
What’s more, reinforcement learning makes up a key part of DeepMind’s deep Q-network (DQN) thanks to its ability to disseminate learning across many workers in the hunt for various things such as attaining superhuman performance, especially in Atari 2600 games.
Despite the perks associated with reinforcement learning frameworks, they consume a considerable amount of time in mastering a goal, are at times unstable and tend to be inflexible.
For these reasons, Google proposed an alternative: an open source reinforcement framework that is based on TensorFlow, which is the company’s machine learning library. The groundbreaking framework is currently available on Github.
In a blog post, Marc G. Bellemare and Pablo Samuel Castro, Google Brain researchers, acknowledged that the recently unveiled platform was inspired by one of the core components in reward-motivated behavior mainly in the brain as well as those reflecting the strong historical links between reinforcement learning research and neuroscience.
They added that it intends to facilitate the type of speculative research that can propel radical discoveries. In conclusion, Pablo and Marc wrote that the release consists a several colabs that help in clarifying the framework ’s use.
The open source reinforcement learning framework was developed with three aspects in mind including reproducibility, stability, and flexibility. It also comprises a compact selection of well-documented code, which is based on the Arcade Learning Environment ( a platform that uses video games to evaluate AI technology).
The framework also features four different machine learning models including a simplified variant of the Rainbow agent, DeepMind’s deep Q-network(DQN), the Implicit Quantile Network, and C51.
As far as reproducibility is concerned, the code is supplied with full training data and test coverage, specifically in Python pickle and JSON formats, spread out across all the 60 games backed by the Arcade Learning Environment. It also adheres to the best practices, particularly on standardizing the empirical evaluations’ results.
Aside from the launch of the open source reinforcement framework, Google is also unveiling a website that aids developers in rapidly visualizing training runs for different agents. Other than that, the Mountain View-based company is also expected to avail TensorFlow event files, raw statistics logs and trained models, primarily for plotting with TensorBoard, its visualization tools suite for TensorFlow programs.
Castro and Bellemare also wrote that their hope is that both the framework’s ease-of-use and flexibility would help in empowering researchers in trying out new concepts irrespective of whether they are radical or incremental.
They also confirmed that they are currently leveraging the reinforcement learning framework in research and have found out that it gives them the desired flexibility needed for iterating over multiple ideas quickly.
For that reason, the Google Brain team researchers expressed their excitement to see how the larger community would take advantage of the framework.