AI Article in EE Times

Posted in artificial intelligence at 10:18 am by JamesNeel

Here. Basically, they’re looking to make a computer-controlled avatar that could pass the Turing test. While there’s not a lot of technical depth there, from a CR perspective there’s some insights that can be made from the following excerpts.

Mimicking the behavior of a human-controlled avatar in a virtual world like Second Life is possible, according to Bringsjord, if you craft the necessary algorithms carefully and run them on the world’s fastest supercomputer. Bringsjord’s synthetic-character software runs on the supercomputers at CCNI, which together provide more than 100 teraflops, including a massively parallel IBM Blue Gene supercomputer (the title-holder to world’s fastest supercomputer), a Linux cluster-supercomputer, and an Advanced Micro Devices Opteron processor-based cluster supercomputer.

Rascals is based on a core theorem proving engine that deduces results (proves theorems) about the world after pattern-matching its current situation against its knowledge base. Each proven theorem then initiates a response by virtue of having a synthetic character speak and/or move in the virtual world.

“Upon analysis, anything that our synthetic character says or does, is the result of a theorem being proven by the system,” said Bringsjord. “So far, theorem provers have only been used in toy-problems. We are scaling that up to enough knowledge for a synthetic character, which requires a very fast supercomputer.”

Bringsjord’s research group recently passed a milestone by programming a synthetic character to understand a “false belief.” For instance, to create a false belief you could hide a stuffed bear in a cabinet in front of a child and an adult, and then when the adult leaves the room, move the bear to a closet while the child is still watching. Here, the child should know that the adult now has a false belief–that the bear is still in the cabinet.

In general, we (the CR community) will cannot assume our radios will have their own supercomputers (at least for a decade if not longer).

  • Our radios’ run-time adaptations will need to be determined by far simpler algorithms than what is being currently explored in the AI community
  • This could be supplemented by case-based reasoning to choose algorithms
  • Attempts to automatically define solutions for completely novel problems should likely only be handled during off-line processing (to establish cases for online processing).
  • Even then, it will likely be necessary to restrict our reasoning algorithms to scenarios analogous to the “toy-problems” alluded to in the preceding. In practice this means highly restricted reasoning domains.
  • Since we shouldn’t be trying to make the 100% solution in our first generation of CR (or even the second or third generations), we can and should be constantly looking for ways to “cheat”. The AI community has to infer what people are thinking. For inter-radio reasoning, one radio could just ask / download what the other radio knows.

Comments are closed.