TL;DR:
Short week, since I took Monday off to see a visiting friend and went to Bainbridge Island with family for July 4th. (Bloedel Gardens pictured) Wrote a lot of latex, formalizing the logic behind fancy observation goal recognition. It’s made great progress. Didn’t do much coding, except to re-do the java prototype to fix sets. I’m trying in earnest to get this published.
This Week:
- Goal Recognition:
- Wrote a ‘Hero Narrative’ pitch for complex observations in goal recognition, which will serve as an argument outline for publication
- Wrote a good chunk of the formalization for complex observation sequences, intended for publication.
- Fixed my java prototype parser so that sets are necessarily ordered.
- Language PDDL: I’m trying to read Attention, Intentions, and the Structure of Discourse, and it’s very… dense. Considering linguistics is all about the study of language, you’d think their field would have better terminologies.
- Broader Studies: I read this paper, titled Biologically plausible deep learning — But how far can we go with shallow networks? It’s not my area, so I didn’t focus on details or much on results. I’d never properly realized why neural networks weren’t like real brain neural networks. I knew that real neurons fire in bursts, as opposed to continuous values, but I didn’t realize that human brains use an entirely different learning algorithm. (We don’t know what it is, but it certainly isn’t backpropogation.) The paper just ran tests on a few combinations of other learning algorithms that might be what human brains use. They carefully said “biologically plausible”, not “biologically likely”. The study wasn’t really to understand brains, but to explore new learning algorithms using biology as a guide towards locally efficient models.
- GRE: Did not do anything. :/
Next Week:
Goal Recognition:
- Continue with the formalization, write introduction first draft
- Implement my parser atop the Ramirez
- Be on the lookout for how to do observed actions with missing parameters, so as to create mutex sets for the possible instantiated actions
Language PDDL: Read more of that paper.
Apartment Hunting: Matt and I gotta live somewhere when we get back to school.
A Plan for Publication
My Plan Recognition project is turning out well. My advisor says this work is worth a conference submission, so we’re shooting for AAAI 2020 which is due in late August. He also says that this would be a nice easy Honors Thesis, if I wanted to use it for that. There is something very tempting about having it all done…
That leaves two months and a finite number of steps towards August 30th.
- Write code to automatically generate the plan recognition domain from complex observations
- Write introduction and descriptive bits of paper
- Finish the formalization part of the paper, and get feedback
- Put together a system for generating evaluation data and running the evaluation
- Something to make optimal plans for benchmark planning problems
- A plan “runner” to get a trace of the plan (each intermediate state)
- Select subsets of those plans/traces, varying how much is observed, how much is in an uncertain order, and who is partially obscured. Vary also if using action observations, state observations, or a mix.
- Run evaluation
- My own formulation for complex plan recognition
- Run the original formulation a bunch of times to find the probability of correctness.
- Code to turn my partial order stuff into all combinations of total orderings
- Time to run it all
- Analyze and write up results
- Prepare for submission
- Submit