Today is a big day for TR. Congratulations to TR!
Adaptive Training uses machine learning and the power of an unprecedented data set to ensure you get the right workout, every time.
I believe there is a lot of things to talk about on this topic.
- What kind of neural network used in the current TR plan?
- Is this Supervised learning or not?
- What is the strategy for anomaly detection?
I believe Machine Learning is quite new in the cycling industry, so there is a ton of opportunity for us to learn new things. This I believe is more fun than just raising my FTP.
What’s more important, how can I verified and validate the adaptive plan is more effective within a given time frame from an individual perspective?
What do you think about this topic?
Do not be shy. Any ideas about ML in cycling are welcome. Books, code, papers, podcasts, ideas to test, your personal experience are all welcome to share.
On a related note, I’m wondering what physiological knowledge was used in this process, if at all. Was this a pure big data/ML project? If not, how was physiology used to guide the process.
The challenge is not everyone agrees on what is best to proceed. Long term…this is definitely a good approach. Short term…expect a lot of bumps.
The unfortunate answer is you cannot. Same with the question on an individual level of would you get larger gains doing traditional base versus sweet spot vs polarized vs XX type of training or volume. The issue is that there is no way to run an A/B comparison on yourself.
On a large data set, TR could see things like:
- People who use / accept the adaptive training recommendations improve their compliance
- People who use / accept the adaptive training recommendations improve their FTP (yes, I realize this isn’t necessarily the gold standard in defining improvement) on average by XX% more than people who don’t
Or whatever metrics they define to mean “success” for adaptive training. And then tweak the models / learnings / etc. accordingly to try and improve these metrics.
What I would like to see TR go into more detail on is: what are the goals / objectives / etc. they are trying to optimize the adaptive training model for? And what was the thinking behind these, and what possible downsides are they worried about with these vs. other objectives / metrics
Why assume a neural network over other model implementations? Could also be using ensemble modeling.
Does anyone know where the benchmark 10 score comes from? Is it YOUR 10? Is it your age groups 10? Is it the whole of Trainer Road’s 10 or Peter Segan’s 10?
The base of your training is the volume and frequency of training. It supports everything else. If you’re not completing your workouts then volume, frequency and compliance to any plan drops as motivation falters.
Just this one aspect will improve your training results.