Elon Musk builds ‘gym’ to train AIs
Thu 28 Apr 2016
Tesla and SpaceX owner Elon Musk, is now establishing an artificial intelligence ‘gym’ for developers to train their AI systems through a series of games and challenges.
Under Musk’s $1 billion non-profit OpenAI initiative, the gym is now available as open source code and provides ‘environments’ for developers to test their new AI bots. This includes 59 Atari games including the popular Alien, Air Raid, Asteroids and Pac-Man.
The results from these test environments will be compiled by OpenAI into a list of the top performing systems. Instead of ranking the AIs by high scores, the systems will be judged on versatility.
“It’s not just about maximising score; it’s about finding solutions which will generalise well,” the OpenAI guide explained. “Solutions which involve task-specific hard-coding or otherwise don’t reveal interesting characteristics of learning algorithms are unlikely to pass review.”
The free open source toolkit also includes board games and physics simulators, as well as many traditional AI training situations such as driving a car uphill and resting a bar on a moving cart.
This model is based on Reinforcement Learning (RL), which studies the performance of AI exposed to complex and uncertain environments in response to the promise of reward. The technique was used to develop Google DeepMind’s AlphaGo AI which defeated Go world champion Lee Se-Dol in March.
OpenAI claims that research in this area has slowed recently due to a lack of standardisation and sufficient benchmarks. The gym setup hopes to solve both of these roadblocks.
Announcing the new platform, co-founder and CTO Greg Brockman and researcher John Schulman explained that the gym was originally built for their own research needs, but ‘hope it will be just as useful for the broader community.’
The OpenAI gym, currently in beta, is initially available within Python, with plans to update for any language and expand the selection of environments in the near future.