Ancient Greek philosophers mused on nature and nurture and how they inform human behaviour, and social scientists continue this debate today. A new DeepMind scalable environment simulator takes a digital approach to the question, enabling the examination of environmental factors on AI agents.
The DeepMind Lab2D platform is designed to create “2D, layered, discrete ‘grid-world’ environments in which pieces akin to chess pieces move around.” It targets the specific needs of multi-agent deep reinforcement learning (RL) researchers, and has been open-sourced.
The DeepMind team chose 2D environments as they are inherently easier to understand than 3D environments, with “very little, if any, loss of expressiveness.” The trim format enables researchers to capture AI problems’ relevant complexity without considering the continuous-time physical environment. Moreover, 2D environments are more comfortable to design and require significantly fewer resources to run. In other words, the researchers can complete intensive calculations and tasks of model and algorithm training on standard CPUs, with advanced GPUs or specialized hardware not required.
To demonstrate the Lab2D system’s support for multiple simultaneous agents interacting in the same environment, the team used an imperfect information game called Running With Scissors.
The researchers explain that agents “must decide how to ‘play rock’ (or paper, or scissors), in addition to deciding that they should do so.” To increase rewards, agents have to identify other agents and what resources they are collecting (e.g., rocks) and secure the resource corresponding to its winning counter-strategy (e.g., paper). There also exists the possibility of agents observing their opponent’s policy development and taking countermeasures, which induces the potential for “feinting strategies” that could not be easily captured in a classical matrix game formulation.
The paper shows how 2D environment simulations require much fewer resources compared with 3D environments. The researchers had two random agents compete in Running With Scissors on a full RGB size of 80x80px, training with an average of 250,000 frames per sec on a single core Intel Xeon CPU at 3.7GHz, and the cost of running the simulation was nearly negligible.
DeepMind stresses that progress toward artificial general intelligence will require “robust simulation platforms to enable in silico exploration of agent learning, skill acquisition, and careful measurement.” To this end, it has open-sourced Lab2D to facilitate the design of multi-agent learning and intelligence tests.
The paper DeepMind Lab2D is on arXiv, and the code is available on the project GitHub.
Analyst: Robert Tian | Editor: Michael Sarazen; Fangyu Cai
This report offers a look at how China has leveraged artificial intelligence technologies in the battle against COVID-19. It is also available on Amazon Kindle. Along with this report, we also introduced a database covering additional 1428 artificial intelligence solutions from 12 pandemic scenarios.
Click here to find more reports from us.
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.