Monday, June 16, 2008

The macro on decision logic

Decisions we take are either intuitive or based on a calculated risk. Decision takers include you and me, as also economists/bankers, politicians and gamblers (to define the spectrum of decisions that this post talks about).

Economists have their intricate mathematical models, politicians - their gauging of the collective sentiment and gamblers, their gut. All of these decision-making devices, fuzzy, neural, artificial, intuitive, etc are used on a daily basis to predict the future. That said, the world's population would all be paupers or trillionaires if any of these could boast consistency.

Given that such is not the case, i.e. there's only such a percentage of accuracy any decision logic machinery can claim, it's time to start looking at models that are organic - Letting reality and recallable (recent) history be the major inputs to your guesstimating. As an example, if the roulette ball lands on the 00 six times in a row, and you're unsure where to bet next ("hmm....even or odd?"), just go with the 00. Chances are the house is playing dirty. The Mahabharata would've been a shade less vengeful had Yudi figured this out before betting the kitchen sink.

Of course, this isn't revolutionary. It's part of any adaptive system in use today and all connectionist models rely on history (wiki this stuff). Here's another element to help focus the predictions - Your best guess for the future. Toss this element into the mix of inputs (i.e. along with what's happened already). An instance of where this might come in handy is with work flow automation at a call-center. If an extraordinary event has happened (an earthquake, discovery of a faulty part in the computers you sell, etc), it's safe to surmise that the future volume of calls will need more than the past week's call-volume as an input to forecast what's coming.

The last decision-making factor is the cornerstone of AI. Is there any information that can be derived from the interaction of your inputs? An instance of this is a loan officer assigning points to a prospective loanee based on standard criteria like their age (assume age directly proportional to points) and whether they rent or mortgage a house (assume renting gets higher points than mortgaging). This would mean that an elderly renter would be a better bet for the loan officer than a young house-owner. In the real world though, this is counter-intuitive, but you wouldn't get that from the discrete point-assigning system. And this is only two dimensional, the logical progression would be to extend this to n dimensions whose interactions can paint a picture.

My point though is to be able to spot patterns when theoretically none should exist - and to be an intelligent human, the macro-decider on when a decision needs to be made via an intelligent machine and when only in collaboration with one.

No comments: