The DARPA has announced that it will be using a Generalized Integrated Learning Architecture (GILA) system from Lockheed Martin to manage crowded airspace. It is intended to help the Air Force to cope with increased air traffic, especially as unmanned aerial vehicles and other airborne weapons get more common.

In effect, the DARPA is planning to hand over the control of missiles to an AI system which is able to learn and reason on its own. Brilliant! According to DARPA, the software will combine reasoning systems, general knowledge and by asking what-if questions.

“The integrated learner also will have explicit learning goals, keep track of what it does not know, what it needs to know, as well as track and reason about its uncertainties. The software will attempt to figure things out, as well as tolerate errors and missing information by using whatever information or reasoning is available. Integrated learning software must be able to manipulate many different forms of information and even trade off different types of information and reasoning, as well as interact with humans to fill in information gaps.”

Let’s just hope it does not learn from the wrong examples, or find out that it can work much more efficiently when it completely shuts off human influence from its controls. Many researchers believe that such reasoning systems are not unlikely to come to wrong solutions that result for example because of poor problem domain description (e.g., the goal might be to minimise the probability of a crash at a given future date. A solution is to crash all planes now, then there are no planes in the future, i.e., the probability of a crash will be zero …). Also, a learner that learns from positive and negative examples is prone to misclassifications. Therefore I’ll rise the AI Panic Level by +1%.

I believe that it might be hard to spot errors in the rule base and inconsistencies in the examples used to train the software. A system able to predict outcomes based on examples has to have some kind of bias, a restriction to possible world states. This bias must be chosen carefully to avoid empty solution sets (too strict restrictions) or full sets (too lax restrictions, the system is unable to make useful predictions).

As long as this system is used in training and double-checking the flight operators decisions only, there is no direct danger. But if DARPA decides to go live with this system (which might never happen though), it is of course critical that no errors occur. This depends crucially on the proper set-up of the background knowledge and underlying rule bases. Lets hope they get it right and make no precipitous moves to replace those error-prone humans!

Stumble it!   Written by Robin

Popularity: 11%