Artificial Intelligence


I'm not committed to the idea that consciousness is an epiphenomenon of complexity. I didn't get into AI for the Grand Challenge (as it is called in the field) of creating a sentient machine. I got into it because it's fun to make computers to cool stuff.

State of AI
Successes that everyone now takes for granted: The hard parts:
Voice recognition and synthesis (cell phones) Common sense reasoning & the frame problem
Expert systems (retirement planning on-line) Natural language understanding
Machine vision in constrained environments (OCR, automated pap smear screening) Machine vision in unconstrained environments (recognizing everyday objects and navigating among them)
Machine learning & pattern recognition (e.g. fraud detection)  
Planning (mission and logistics planning)  

Here are some AI projects that I've worked on. Most of these have been published here and there, but I'm too lazy to type in the citations.

Project Impact Advisor A schedule critic that I built for the Defense Systems Management College. It performs resource-constrained scheduling, then identifies risk areas and provides a narrative assessment.
Thallium Diagnostic Workstation Built for the USAF School of Aerospace Medicine to screen fliers for minimal coronary artery disease. Uses machine vision to extract features from digitized thallium scintigraphy (radionucleotide heart scans), machine learning to create rules for recognizing narrowed arteries (i.e. predicting the findings of cardiac cath), and an expert system to apply the rules to specific cases.
Knowledge Acqisition Techniques for Image Analysts (KATIE) For the CIA Office of Research and Development. An assessment of a dozen different pattern recognition techniques' ability to detect a characteristic EEG response that occurs when someone has to stop and think about something.
Entry/Exit Identity Resolution Used machine learning to derive rules for determining if two border crossing records belong to the same individual, despite lack of exact match on all data fields. The algorithm was trained against records where identities could be confirmed using USVisit fingerprint data.
Patent-In Effectively a chart parser, this system is used by hundreds of pharmaceutical companies to determine the protein sequence that will be expressed by genomic or ribosomal DNA/RNA in the presence of specified coding features. Used to create patent applications for the US, Europe and Japan.
Fleet Command Center Battle Management Program (FCCBPM) A DARPA-sponsored expert system to determine how best to reconfigure the fleet to make up for a sunk/damaged ship. This is the only system listed here where I didn't do the design and development. Provided advice to DARPA/CINCPACFLT. Am currently helping Nils Nilsson place FCCBMP in historical context for the book on AI that he's writing.
Qualitative Reasoning System (QRS) For a commercial client, used model-based reasoning to perform fault diagnosis in complex electromechanical systems (Apache helicopters) with minimal time and effort.

Hope to be adding to the list over the next couple of years.