Manan-Akaash-Lucas 06-05-20

I’m going to keep working on the planning model for now

Try to investigate something more exotic than a vanilla LSTM
- Don’t have a very clear idea of what that would be
- Just look into time series prediction model
- Propose a better model, and try it
- LSTM w/Attention
  - Lucas started it, but doesn’t think that the evolutionary mechanism part of it was best.
  - Probably just to output an attention component
- Fourier transform data
  - Never actually removed the preceding 10 datapoints, could be improved
Shifting towards models that are less data intensive - techniques that need less data 202006051627
Bayesian Optimization could be a cool thing to explore for hyper parameter training 202006051628
Bayesian optimization could be a good framework for neural networks as well
- Propagate uncertainty through the net, get a good spread for points that you’re not really sure about
Auto ML stuff 202006051630
- Tries different instances of models for prediction problems that we’re using, and different hyperparameter for those instances
Take the next week to explore 2-3 of the things that we just spoke about, can measure by comparing RMSE, plots, etc.
- Then continue it, delve into it.
- At the end of the week, doesn’t seem like it’s going anywhere, then I can switch to something
We want to have uniformity in all the planning models that we are testing, to make sure they are uniformly comparable
Stay away from: feeding in information from other points, so that we have a shrinking effect
Make sure that the points vector that’s being explicitly fed in to the model (points vector at time t), is the only points vector that’s being fed in, not points vectors corresponding to different time predictors
Look into approaches for the baseline energy calculation, so we can easily separate the energy from before the experiment starts, and during the experiment, and how we can structure our planning model to decompose the impact of the points on the baseline energy effectively.

uid: 202006051619 tags: #raise #meetings

Date

February 22, 2023

Up next

Collections In Javascript Objects are good for static information, something like this: Not good for dynamic information though. They’re useful if you’re not mutating a

Previously

General updates Lucas brought planning stuff to a close over the last few weeks Looked into Auto-ML stuff Singapore officers are still pretty closed, not fertile