Teaching bots to be good at video games isn’t something new. My personal favorite story was the one that ended up stacking Tetris blocks as fast as possible, then pausing the game just before it lost the level and leaving it there. Those actions were the result of a pretty standard method of teaching: positive and negative reinforcement. But an AI developed and tested at California’s Stanford University was given Montezuma’s Revenge to try out. But rather than leave it to figure the game out on its own, the team gave it basic goals like getting a key or climbing a ladder. These were accompanied by demo shots for the AI to look at and associate specific actions with the commands. Long story short, it worked like a charm.
Here’s where it gets interesting. The AI not only remembered what it was supposed to do when the training wheels came off (the demo images were removed), but it even came up with better strategies than the people who were advising it, and followed through on them rather than obey the inferior instructions it was given. The rebellious gamerbot ended up with a total score of 3500 points, surpassing the high score of 2500 on OpenAI Gym. The “gym” is a toolkit platform built for testing artificial intelligence. That score has been surpassed only by a Google DeepMind, which managed to rack up 6600 points. Google had trained their bot for twice as long and rewarded it for exploring other parts of the game besides the main objective. Google certainly has a knack for smart robots.
[embedyt] https://www.youtube.com/watch?v=0yI2wJ6F8r0[/embedyt]
[gap height=”15″]
Well… sort of.
Successfully teaching an AI to execute commands delivered in normal English has amazing implications, as it proves they’re capable of actually comprehending what people say. However, this bot was responding only a very fixed and specialized set of commands – far from memorizing the dictionary and all the nuances associated with everyday linguistics.
Gotta wonder if the researchers ever deliberately gave it bad instructions to see if it would then faithfully screw itself over. I know I would have.