Logical Inferences Per Second (LIPS) vs. Horsepower

Of course we're a little jealous of those developers who get to develop those fun and nifty IPhone apps! Perhaps we're just a tad bit curious too. But for the moment we are absolutely in the grips of a very different kind of software development.

It is a tricky distinction to make when we talk about the computer in terms of solving the problem as opposed to being part of a solution to problem. I guess in some ways its all a matter of perspective.

So if we can put it another way, we are currently involved in developing software that uses the process of inference to reach conclusions that when applied or understood represent the solution to some (normally nasty) problem or set of problems. This inference process usually takes the form of logical deduction, induction, or abduction. We want to contrast the process of inference to that of computation. So instead of focusing on the work required for a computation we focus on the work required for an inference. In particular we look at how many logical inferences an agent might use to solve some problem. Normally we are concerned with how many inferences an agent might use in the worst case scenario to solve the problem. Once we have some idea of how many inferences that is then we know how big the problem space is and how much computing power we could potentially need. For instance, in our 19 emails project, the worst case scenario for our agent was 19! * 3000 inferences. From a hardware perspective if I have a quad core processor and each processor is clocked at 1.8 GHZ, what are we talking about in terms of LIPS? That is 1.8 GHZ = How many LIPS?

Much like horsepower the metric of LIPS is from days gone by. The researchers at ICOT and the Fifth Generation Project used it as a way to measure system performance. ICOT had a target of 100 MLIPS (million LIPS) to 1 GLIPS (billion LIPS). At the time 1 Logical Inference Per Second took approximately 100 machine instructions. But who knows what machine instructions are these days? (I heard recently that they are similar to lines of Javascript.) Remember parallelism even massive parallelism was at the heart of ICOT and the Fifth Generation project. They produced answers in that project for questions that we are really just now asking today. In that project LIPS was the basic unit of measure for system performance. The ghosts of ICOT have presented me and Tracey with a very seductive argument and we are currently on some DaVinci-Code-Indiana-Jones-Raider-of-the-Lost-Ark adventure to discover the true path to massive parallelism by looking at the stones that were turned over by ICOT and the stones that weren't. At one point in time horsepower was a very hands on metric. The common man knew how many horses it took to plow a plot of land, pull up a tree stump, or get a stage coach to town in a day. At some point when steam engines are introduced horsepower was the metric to use to compare how much work a steam engine could get done. I'm still a little baffled why 400 horsepower is still used as the metric for my little SUV. I have no idea how many horses it takes to pull up a tree trunk or plow a field. I am sure it has meaning for somebody somewhere that has to get their hands dirty. LIPS has a similar function for those of us stuck with the totally futile task of trying to solve AI-complete, or AI-hard problems. Somebody manages to think of our cute little SUV in terms of 400 horsepower and Tracey and I have to think of duo, quad, and eight core * eight hardware thread processors in terms of LIPS. There is a very important relationship between LIPS and our ability to cope with massive parallelism. Stay tuned ...