Karen presents her work at SRI on a Personal Learning Assistant (PAL). This learning differs from Peter’s cognitive presentation in the sense that she is not recreating a “robot” that aims to think and act like a human but rather an assistant that can learn specifics tasks to automate.
The learning approach consists of 3 steps:
The demonstration trace is turned into generalized expressions that look like: A(-[a,b,c], -d, -e) meaning that the operation A was performed with a series of actors. The system has to be enhanced to reach out to external systems for additional inferences such as ” is how to contact the assistant of Karen Myers”. The all of those atomic operations are strung together in some kind of scripting language.
So far it resembles the scripting capabilities you can find in modern editors. It is one of the latest additions of Model Builder, before my departure from FICO. Although we did not intend any intelligence there, just the ability to replay common operations. This is obviously something that modelers are used to since their work is very repetitive.
In Karen’s LAPDOG (Learning Assistant Procedures from Demonstration, Observation and Generalization) project, an interesting aspect is that those procedures can be triggered automatically “while you are away” for example instead of manually as it typically the case (or as part of a modeling workflow).
Karen presented an interesting use case: CPOF Command Post of the Future. It is a collaborative system for sharing and visualizing data. This is where the “while you are away” capability makes a lot of sense in surveillance interactions. It begs the question whether wars could turn into AI wars in the future to avoid human casualties… Just a thought…
What I missed in the presentation was the linkage to rules. I think this is useful of course but somehow I was misled to think that the tasks would turn into Business Rules. My mistake.
Learn more about Decision Management and Sparkling Logic’s SMARTS™ Data-Powered Decision Manager