ONCE AGAIN artificial intelligence (AI) tech has been trained to trounce humans at games, this time making a mockery of them in competitive computer game Dota 2.
Boffins at research lab OpenAI, a lab backed by Elon Musk despite his AI fears, created a team of AI agents that can beat the top one per cent of amateur Dota 2 players, evolving their work on beating such gamers in 1v1 matches to tackling humans in 5v5 games.
The researchers used reinforcement learning, a rather contemporary machine learning technique which involves training AI agents to learn through trial and error with them being ‘rewarded’ every time the did a good thing in the game such as killing an enemy.
The agents played the equivalent of 180 years of Dota 2 gaming every day thanks to accelerating the games time to a heady rate.
However, limitations are in place to only being able to use five out of the 115 hero characters available, which strips away a great deal of the complexity of the game for the AI agents, while other complex features of high-level human play, like setting wards, were also disabled.
So OpenAI’s AI agents are not likely to be able to beat humans in a full-fat version of Dota 2 quite yet, but in a limited state, they can beat amateur players. We suspect professional Dota 2 nerds who dedicate hundreds to thousands of hours to the game would steamroll the smart tech, though.
The tech demonstrates that reinforcement learning can be used to make sense out of masses of data and that AI lessons learnt by OpenAI have the potential to be applied to other things.
“Relative to previous AI milestones like Chess or Go, complex video games start to capture the messiness and continuous nature of the real world. The hope is that systems which solve complex video games will be highly general, with applications outside of games,” said the researchers.
However, given the amount of data needed and the simulated time taken to train the AI agents, the research also shows that super smart AIs are still some way out, certainly with the use of reinforcement learning.
But AI research is evolving at a heady pace with the likes of Google finding new ways to add smarts to algorithms and systems, so don’t go expecting AI to remain a bit dumb for too long. µ
Source : Inquirer