DeepMind recently open-sourced a software system called Lab2D. Lab2D is designed to support the creation of 2D environments for artificial intelligence and machine learning research.
The DeepMind team states that 2D environments are naturally more straightforward to understand than 3D environments at a little loss of expressiveness. The researchers say that even a simple game Pong, which consists of three moving rectangles on a black background, can capture something primary about table tennis’s real game. This abstraction superficially makes it simpler to grasp the nature of problems and concepts in artificial intelligence.
The researchers assert that deep complexity accompanying numerous dimensions can be studied easily in 2D as in 3D. Besides, 2D environments are significantly less resource-intensive to operate and typically do not need specialized hardware, such as GPUs, to attain acceptable performance. 2D environments have been successfully used to study various diverse problems such as social complexity, navigation, imperfect information, abstract reasoning, exploration, etc.
Lab2D facilitates the creation of 2D, layered, discrete “grid-world” environments in which each piece (similar to chess pieces) moves around. This system is specially tailored for multi-agent reinforcement learning. It supports multiple players interacting in the corresponding environment simultaneously. These players can be both human and computer-controlled. Each player can have a custom view of the world that shows certain information. A global perspective that is likely to be hidden from the players can be set up and include specific details. This can be used for imperfect information games where players don’t share standard information and for human behavioral tests where the experimenter can observe the environment’s global state as the episode progresses.
Lab2D provides many mechanisms for revealing internal environment information. One of the simplest forms is observations that empower researchers to add specific information from each step’s environment. The second way is through the events. These are not bound to time levels but are rather triggered under particular conditions. Ultimately, the properties API offers a way to read and write the environment’s parameters.
Artificial intelligence research based on reinforcement learning is beginning to evolve as a discipline. Research workflows spend considerable time authoring game environments and intelligence tests, adding analytic methods, and so on. These activities are critical to the enterprise’s success. However, they are not as simple and straightforward to extend as they ought to be. The team states that DeepMind Lab2D is a step toward robust simulation platforms that might facilitate learning, skill acquisition, and AI systems measurement at scale.
Lab2D generalizes and spreads a popular internal system at DeepMind, which supports a broad range of research projects. It was wildly popular for multi-agent research involving workflows with significant environment-side repetition. The team found that DeepMind Lab2D facilitates researcher creativity in the plan of learning environments and intelligence tests. They are eager to witness what the research community uses it to build in the future.