Researchers at the OpenAI institute are using the original Sonic the Hedgehog video game to teach AI how to think about previously unseen, complex scenarios and overcome them. Surely it’s only a matter of time until the blue blur helps computers control the world then, right?
Well, apparently not, according to the researchers in the video below, but it’s certainly an interesting experiment. The OpenAI research team, which has previously demonstrated how AI can master P2P video games such as Dota, is currently turning its attention to the 16-bit Sonic series in a new open contest it is running.
But, while past AI experiments have focused on reaching “an extreme mastery” of a video game, the intention with Sonic is to “focus on reaching a competent level [of play] really quickly, in the same way as a human would.” So instead of simply understanding a set of rules and being challenged to compete with others (or to complete a series of tasks that involve repetitive action), here the AI is being challenged to navigate a complex level map, filled with traps, that it has never seen before.
The real task however, is to test whether an AI will ‘remember’ how to overcome individual traps (say, spikes or a loop-de-loop) and apply the solution it creates (to jump over them, or to take a running start to gain enough inertia to clear) across all levels it attempts (from Green Hill Zone to Spring Yard Zone and beyond). This is what’s known as ‘transfer learning’, and can mark a significant step forward for AI development – at least, for those built for playing video games.
Watch the video below by Mashable, which goes into some extra detail on the nitty gritty. It’s super-fascinating – especially the part where the researchers start talking about how AIs under pressure had the potential to glitch the game and use boundary breaks in Sonic 1 to dart through chunks of a stage.
The OpenAI Retro Contest for Sonic the Hedgehog runs until June 5, so if you’re particularly techy and know how to code, why not give it a bash?