This week a deep learning company that Microsoft purchased this January revealed that a computer earned the highest score ever achieved on Ms. Pac-Man.
On the surface, this victory seems pretty silly. Who cares about a little video game? But what it represents has far-reaching implications for the field of Artificial Intelligence. The company taught a computer to break down the game into tiny segments that would yield rewards if the segment was successful. Then the controllers of all the tiny segments sent their recommendations for game play to a master processor, which weighed each move possibility with other options to create the best path for the bow-wearing yellow orb. I love this article that shows how many things we use every day are actually powered by AI. And remember the experiment I did with scheduling with an AI assistant named Amy?
I’ve been holding off on talking about Artificial Intelligence because I wanted to wait until I understood it completely before trying to explain it to you.
Yeah, that ain’t gonna happen.
I’m just not smart enough to figure out how smart computers are and how smart they’re going to get. But the advancements even this week are so interesting that it’s time we talk about them.
Teaching AI to Make Deals
Researchers at Facebook gave pairs of humans ten objects (balls, bats and books) and asked them to divide them up between them. Then they took the results of those human negotiations and fed them into a computer. Researchers then asked the computer to practice with itself to improve its negotiating skills. Each “player” had individual goals for the negotiations and could earn points for better negotiations.
- Computers Can Negotiate as Well as Humans
After the computer played with itself for a while (that’s not a double entendre until computers pass the Turing Test), the researchers asked the bots to negotiate with humans, and they ended up negotiating better deals about half the time.
- Computers Can Learn to Bluff
Some reporters are saying that the AI learned to lie during the negotiations, but I think of it as more of a bluff. The computers bots learned that if they pretended to value one of the 10 items they were haggling over, they could later “give in” on the object they didn’t care about in order to make a better deal. Do you consider that lying? Maybe it’s just an #alternativefact?
- Computers Can Make Up Their Own Language
Here’s what is perhaps the most unnerving of the revelations. Researchers realized that when the computer was learning from itself, the bots used a non-human language to communicate, leading the researchers to reprogram the computers to make sure they only negotiated in English. Read this article from The Atlantic for a complete examination of this point, but the fact that the computer created a more efficient language to talk behind researchers’ backs to learn is… wow. That’s just freaky.
- Computers Can Move Beyond Pre-Programmed Responses
When the bots were deep in negotiation, they could construct sentences beyond the response phrases they had been trained with.
- Many Humans Didn’t Know They Were Negotiating with Bots
They thought they were talking with other people… stubborn ones at that.
So what do you think about AI? How are you using it? Anyone working with Messenger Chatbots or other AI tools? Spill the beans!