Hi, I’m a real beginner with deep learning and I’ve only messed around with it a small amount. So I’ve seen other teams out here with full scouting applications with predictive deep learning algorithms to predict match results. A lot of this seems completely overwhelming so I was wondering (maybe this will be helpful to others looking as well) if you guys have seen any good resources for getting into deep learning in an FRC Scouting environment? Thanks so much!
Deep learning is a niche that takes a good amount of attention to grasp. The most important thing for a neural network is the data, which is, in my opinion, where the problem is in the FRC environment. While you can collect performance data for each match, it is hard to narrow robot characteristics down to a data point. Models need good, solid data. The more info the better.
That being said, I wish you luck in your ventures. You have to start of with the basics and work your way up. I personally took Andrew Ng’s ML course on Coursera and then loved Practical Deep Learning for Coders from fast.ai.
You can find more courses from articles like this
Check out part 1 of Fastai’s MOOC. I developed all the predicted match scores/robot stats features for my team’s (Team 2102) scouting application last year using key principles from this course. I suggest that you’ve got a decent background in high school math, Python, and a willingness to learn/struggle with new concepts as you go into the course as it took me a few times to go through all the lessons to really understand what I was doing. I’m actually working on a write up right now on how we developed our scouting application RoboRecon which includes all the ML/Deep Learning work. PM me if you’ve got more questions.
There’s really not enough data out there to develop comprehensive predictions. I’d love to be proven wrong, but so far every scouting or prediction system I’ve seen that uses any sort of machine learning either produces highly inaccurate data, makes a lot of assumptions (producing pretty but generally inaccurate data), or barely outperforms a simple OPR or win-loss rate comparison. Neural networks need a ton of data, and that data just isn’t out there.
What do you mean by scouting? We have clipboards and printed sheets to scout all teams from auto to endgame. It’s not 100% accurate as there can be a few misunderstandings and our scouters may not understand them. We collect these and look at them at the hotels, in the stands and the pits. We’re usually forgiving on penalties but constant reoccurrence well be significant. But that’s the way our scouting goes.
Sent from my Galaxy S8 using Tapatalk
I think OP is talking about using the data you collected (by app, paper, scantron, whatever) to make match predictions/give picklists using deep learning.
We use tablets to get scouting data with an app, that later gets compiled into an excel spreadsheet. Our mentor works with statistics quite a bit (CD username “Whatever”) and he made an extremely accurate FTC calculator a few years ago using algorithms which he then implemented last year for FRC. I’m sure if you have questions he could help!
I have used the following data sources in the past to build predictive models. Both sources update and publish weekly on during the FIRST Robotics Season.
Caleb Sykes posts white papers and past season statistics, as well as a scouting database. Ether also publishes raw data and analysis.
Start with the White Papers published by Caleb: https://www.chiefdelphi.com/media/papers/3315
Caleb’s scouting data base:
Ether’s published raw data.