What are some impressive algorithms or software in the world of AI?
I have always loved the idea of AI and evolutionary algorithms. Unfortunately, as we all know, the field hasn't developed nearly as fast as expected in the early days.
What I am looking for are some examples that have the "wow" factor:
Self-directed learning systems that adapted in unexpected ways.
Game agents that were particularly dynamic and produced unexpected strategies
Symbolic representation systems that actually produced some meaningful and insightful output
Interesting emergent behavior in multiple agent systems.
Let's not get into the semantics of what defines AI. If it looks or sounds like AI, let's hear about it.
I'll go first with a story from 1997.
Dr. Adrian Thompson is trying to use genetic algorithms to create a voice recognition circuit in a FPGA. After a few thousand generations, he succeeds in having the device distinguish between "stop" and "go" voice commands. He examines the structure of the device and finds that some active logic gates are disconnected from the rest of the circuit. When he disables these supposedly useless gates, the circuit stops working...
Edit
Can we try and keep the discussion to techniques/algorithms that produced something impressive? I can google if I want to read about the thousands of AI technologies that are in the early stages but showing promise.
Answer
I built an evolutionary algorithm for retail inventory replenishment in a product targeted at huge plant nurseries (and there are some really big, smart ones -- $200m companies).
It was probably the coolest thing I've ever worked on. Using three years of historical data, it crunched and evolved for a week straight while I was on vacation.
The end results were both positive and bizarre. Actually, I was pretty sure it was broken at first.
The algorithm was ignoring sales from the previous few weeks, giving them a weight of 0 for all indicators (which is at odds with how these guys currently work -- right now they consider the same week in the previous year and also factor in recent trends).
Eventually I realized what was going on. With the indicators the organism had to work with, over time it was more efficient to look at the same part of the previous month and ignore recent trends.
So instead of looking at the last several days, it looked at the same week in the previous month because there were some subtle but steady trends that repeat every 30 days. And they were more reliable than the more volatile day-to-day trends.
And the result was a significant and reproducable improvement in efficiency.
Unfortunately, I was so excited by this that I told the customer about it and they cancelled the project. That first run was extremely promising, but it was hard to sell as proof even though you could crunch almost any data from the last three years and see that the algorithm magically improved efficiency. EA's are not hard, but people find them convoluted at first, and the idea of doing something so arcane was just a little bit too much to swallow.
The big takeaway for me was that if I ever create something that appears a bit too magical, I should hold off on talking about it until I can put together a good presentation. :)