AI planning
I recently published my first piece of academic research, and thought I would Write Something Online about it. Shoutout to my co-author Pascal Bercher. The research process was a lot of fun, but I likely don’t have the constitution to be an academic - too impatient to submit everything I do to a journal, although I hold huge respect for the people who do.
The research was on AI planning. TL;DR AI planning is a form of artificial intelligence that is deterministic and can be trusted with dependable decision-making - it makes decisions based on a description of the world state and a bunch of algorithms/heuristics to find the best path to the desired end state. My analogy is that you would never want ChatGPT flying your plane; but you could trust a (really, really, futuristically good) AI planner.
Researchers in the (fairly niche) field have designed domain-independent planners i.e. engines that will apply an optimised set of algorithms to solve any planning problem. Therefore the hardest part of solving a planning problem is now just describing the world of your particular planning problem accurately. Unfortunately, no-one has spent much time building good parsers/linters for the domain modelling languages - so if you try to describe the domain you are likely to spend an extremely long time getting mysterious errors from the engine without having any way of pinpointing what caused the error.
My contribution was to compile possible syntactical and semantic flaws, build an eval set of flawed domains, and then benchmark each domain modelling assistance software to see how successfully they identify each mistake. We open-sourced the repository and published a shortened version of the paper in The Pacific Rim International Conference on Artificial Intelligence (PRICAI). The full paper is attached below.
Unfortunately due to covid I don’t get to fly out to present at the conference in-person, and instead will be presenting over Zoom! Alas.