A new method for predicting humans' scores for plans

摘要

We describe a computer program that predicts different sorts of people's plan preferences - no matter what problem is being addressed. It does this by "learning" each user's generic planning priorities. The latter take the form of the relationships that people assign between a plan's quality and its scores for ten, always-used plan-evaluation criteria. Hence when we estimate any plan's scores for these criteria, the software predicts this plan's desirability on behalf of any past user, or group of users. Such predictions can be made using either the graphical method described here, or a simulated neural network or Bayesian inference. Moreover, predictions can be assigned confidence limits, which allow the program to deduce whether differences between plans are statistically significant. We demonstrate the accuracy of our software's predictions, and show how it can be reverse engineered to document what it has actually learned about its past users' planning priorities. Finally, we explain how our software uses an innovative graphic communication method to explain why different groups will probably favor certain plans. Hence our software constitutes a precise human-computer interaction system for intelligently addressing the problem of plan choice within wickedly unstructured and socially sensitive environments.