At the Speed of Irrelevance

CHAPTER 9AI for Agencies

ON A COLD FEBRUARY morning in 2021, one of the co-authors of this book received an urgent message from a government contractor executive. She, along with a group of several firms, was trying to respond to an AI-related RFP. She wanted Professor Naqvi to review the proposal before submission. When Professor Naqvi reviewed the RFP, he observed that both the questions and the answers related to the software quality were not applicable to machine learning (ML) systems. The RFP, Professor Naqvi recognized, must have been developed by the staff that was not trained in AI. Professor Naqvi explained to the executive that while the question seems to be for the post development, deployment, and production-related integration stage of ML, testing for machine learning application development is widely different than that of non-learning software. Since machine learning software develops from data, its development process is different. It requires training the learning algorithm. For example, among others, some of the issues in testing are:

• Understanding the features of data and initial testing for performance potential and informativeness of features (for example, calculating entropy) and studying mathematical characteristics of the data;
• Dividing data into development/training, cross-validation data, and testing data;
• Selecting various algorithms for testing and studying their performance;
• Understanding the population dynamics from which data is selected and ...