With the complexity of modern AI systems, they sometimes come up with solutions we don’t expect … when we least expect it. Here’s a great video example of this; an AIBO robot is presented with a problem that seems to have only one possible solution. Can you figure out more than one? The AIBO can.
Placed in a dark room with one lit button that turns on the lights, and one unlit button that opens the door out, the AIBO is given the goal of getting out of the room. It’s too dark for him to see the unlit Open button, so the idea is that he should learn to turn on the lights first. Watch the video to see what happens.
Did you figure it out before watching? The AIBO is part of a research program at the Rutgers Laboratory for Real-Life Reinforcement Learning, and the video was submitted to the AAAI-07 video competition earlier this year (and won for best short video).
These kinds of surprising solutions are often fun, and not that uncommon. I’ve experienced similar things myself, where creatures in a simulation I was developing found a way to get around a programming fault — effectively using their own intelligence to counter a mistake that human error had left in the fabric of their universe.
Links and References
- Rutgers Laboratory for Real-Life Reinforcement Learning
- Video Collection from the AAAI ’07 Video Competition
- AIBO on Wikipedia (picture under FDL license)
- Reinforcement Learning on Wikipedia