Glad someone said it: it doesnāt matter how good the machine learning engine is that chooses your workout, it still requires the slab of meat driving the pedals to execute it, and with consistency over time!
Personally, I get enough entertainment from Xert to keep me putting in the hours. Specifically its the EBC, smart intervals, breakthrough detection, and realtime prediction of time-to-exhaustion/MPA that are things I havent been satisfied with [or even seen?] on other platforms for the same money.
These forums are also quite engaging and entertaining too, even when they end in TR vs Xert vs Outdoor vs Indoor tribalism.
The best and most relevant comment I can give regarding the superiority of ML and neural nets over the 50 year old Bannister model is to direct you again to the slowtwitch thread I linked to earlier.
Best to ignore trolls with an agenda especially repeat offenders.
If you are annoyed by someone (including me ) you can always double-click on their name, change Normal in upper right to Ignore, and you wonāt have to worry about taking the bait.
Here is the thing with ML, you/ and even the programmers will have no idea what model it is working on in short order. And to assume they are not seeding it with a Banister based model is unfounded. In the end the TR problem was simple progression and a single variate workout metric so that many users failed their prescribed workout. The new system is a response to that problem and like any model the proof is in the predictive quality. Complex is not always better. To say the Banister method is antiquated is probably an unfair criticism, all of the training programs out today use some modification of that and the critical power model to make their predictions or to show progress. For your reading pleasure: https://journals.physiology.org/doi/pdf/10.1152/advan.00078.2011 One author is P Skiba who has been in this space a while and still uses modifications of the Banister model. TR uses the Coggan modification to use FTP, NP etcā¦ so as time goes on, training equipment like power meters get cheaper, and more reliable, new metrics, and computing power is cheaper we may see models that use more complex data, but it may be that ML etc eventually come to the same resolution, using the parameters already in the basic models, but you have no way to tell.
ā¦and he was called out by another user for not being open to new ideas:
AC, if you want to keep the interest of readers, donāt come off as a prick know-it-all. You are on the fast track to Hambini here.
I was pretty open with some key concepts in that thread that he either didnāt pick up on or chose to ignore. My understanding is that heās building a training platform based on ML. Not too worried. There have already been a number of other attempts at this in fact over the years. None have made it.
In the end, systems that use ML will have the same problem as the simple Impulse Response model has:
How do you measure training and factor in intensity? - you need to move beyond one value!
How do you measure performance, i.e. whatās the objective the ML is trained on? Using MMP and looking for intervals hasnāt worked well and MPA analysis is now standing head and shoulders above this.
Field data is sparse and starting conditions are not known. This is what makes building a system to automate training prescription an order of magntitude harder since it has to work for everyone, not just those with data that matches what the ML was trained on. This has been our biggest challenge and ML will not help in this respect. In fact, ML may actually makes it harder to deal with.
I highly doubt anyone that comes out with an ML-based platform will share their training data and methods. The ML platform in some ways becomes more proprietary because not even those that offer the service will know how it comes up with the recommended training.
Weāve been pretty open about how Xert works. We havenāt published the math but we go into great length to describe the methods. Itās no different than things like TSS and NP. People donāt care about the mathā¦ They only care if it works.
@CarmenV I can see that you spend a bit of time in the STāS forumā¦ and thatās fine but please donāt bring the ST bitter attitude to this friendly forum. Itās a different vibe thatās all.
Alanās attitude defending his corner was in line with the ST vibe, I will guess some folks in this forum might admire him too, but at the end heās just another interesting opinionā¦ not God.
I stopped reading his paper when he stated: āInterestingly, the 2 athletes who the NN didnāt perform well for are the lowest fitness of the group (VO2max in the low 50ās). Most of the remainder of the group have VO2max numbers in the 65 to 80 range (high level AGers and elites/professionals)ā
ā¦ my opportunity in the TdF is past gone unfortunately.
My question to you, in good faith, is what platform are you using TODAY, that uses NN or that gives real-time feedback (on the bike)?
Iām always open to new training methods and science breakthroughs, but Iām also conscious that my bike have two wheels and the wheel was invent more than a few years ago.
āExternal hand cranks start the engine every time. Electric starters are a marketing gimmick, a solution looking for a problem.ā ā¦Henry Ford, 1912, responding to Cadillacās new Model 30, first automobile with an electric starter.
try this analogy: the model T had an internal combustion engine, the Ferrari still has an internal combustion engine running on gasoline, same basic principle, piston, spark ignitors, and valvesā¦ but just advancements in things added. mostly evolution and that is what likely even ML will do for training, the same basic inputs will still go into the engine, maybe a few additives and maybe when all is said and done who knows it may be the same algorithm, but no one will know, wonāt that suck? Anyhow it has been a fun discussionā¦over and out!
) you can always double-click on their name, change Normal in upper right to Ignore, and you wonāt have to worry about taking the bait. 