I have this complaint with all the training software, but Trainerroad has the best forum to discuss it.
So for the last 10 years, an absolute ton of training data has been generated. Fifteen years ago it was just ‘‘60 min ride, max speed 40 mph, average speed 20 mph, max hr 180, average 140’’ etc. Now there’s so much data, but the training apps are dumb.
When Chad sets out my plan, and says ‘‘do this workout Monday’’, and I do it, there are no adjustments made to future work outs of the plan based on my performance. I mean if I’m supposed to do 60 mins, IF .84, x-intervals etc, and I smash the workout and look like it wasn’t that hard, the plan doesn’t adjust. If I do a three hour Sunday ride rather than my 90 min TR workout, the plan doesn’t adjust … maybe my Sunday bunch ride ends up with a ton of climbing and isnt the recovery ride Chad wants me to do … but somehow my Tuesday ride is still intervals … a human coach may switch Tuesday to an easy day based on Sunday’s effort.
With all the data, why doesn’t the software have more AI type ability to duplicate what a human coach would do? I would like TR to have an expected performance in a workout, based on previous performances and knowing my TSS for the past few days, and then if I’m over performing, it makes things harder, and if I’m not hitting the numbers, it eases me off (throws in a recovery for the next ride rather than intervals or vice versa)
I work in product development, but on the mechanical side and not the software. So as a mechanical engineer, it doesn’t sound that hard … but I’m probably over simplifying it all. If the TR app was more like a human coach, it would surely dominate over all the basic apps that don’t adapt.
Xert tries to do some of the stuff you’re talking about. I went through the trial period and I personally believe it doesn’t do any of it very well (e.g. VERY obvious and unhelpful suggestions about when to train, how much, etc). But YMMV.
Smart training apps are coming. Xert is the most obvious example that is taking steps in that direction. I think Today’s Plan may also have some tools headed in that direction.
Hints from Nate about their data, review, and what to do with it almost suggest they might be headed in that direction as well. I hope so and think it is an area that any future-minded training app must consider.
The ideal is something that I see as quite difficult. We are effectively aiming to replace the review and consideration of a coach with automated data review at a minimum. But we all know that the data is only part of the story. RPE, overall fatigue, life stress, and so many other factors also play into the complex equation.
A real life coach can take part in all aspects of that to form a full picture, and then make informed choices and set direction for the athlete. Boiling even part of this into a “Smart Training App” is something that is far from easy.
Based on comments I have seen, even though Xert has some of this, it falls short in ways when our “use case” doesn’t match their planned usage. It makes assumptions that you are doing workouts in particular ways and if now, the data and suggestions will stray from what should really happen.
All this is to say, I love the idea, but as usual… the devil is in the details. Having automated output is totally reliant on the data (garbage in, garbage out) and the calculations within it.
Sadly, I think the need for a person to learn and apply related training knowledge (or enlist another person to do it) will be a long range need for years to come.
Similar apps and services are being developed and just starting to roll out in the powerlifting/weightlifting realm. Juggernaut just announced theirs if I’m not mistaken.
I would be amazing if something like that could analyze our performance per workout and make adjustments via data or a short questionnaire post ride. I’m pretty new to cycling in general and coming off an unrelated injury. I still lift and there are days, sometimes weeks where I’m still pretty torn up from deadlifts or squats. Having something to adjust on the fly would be amazing for those of us who are doing more than just one sport.
There are so many factors to consider, that you’d likely have to record a lot more data (how many hours did you sleep? how well? any discomfort? HR and HRV? dietary intake and timing). I imagine that, like dual-sided power meters, there’d be a lot of data most people won’t put in and they would be unhappy with the results. Not an impossible task, but certainly tough.
After doing the plans and workouts for a while, I’ve gotten pretty good at assessing where I am and when to modify the workout or the plan.
In a perfect world, the software would look at all the data and spit out the perfect answer for optimized training. In the absence of this perfect solution, the next best thing is software that empowers you to make informed and actionable decisions that lead to a positive training result and it seems that’s what the TR team is most focused on at present (eg: with the launch of the calendar.)
There are SO many other factors (sleep, non cycling workouts, morale, etc) that it’s probably a much safer bet to double down on the areas that empower users rather than trying to create a machine that prescribes workouts.
Until AI is here, I would like to have an ability to tweak the plan beyond shifting workouts across weekdays. I.e. implement 8,9-day weeks for those who need more recovery, as well as have 2-on/1-off variations of the plans.
Personally, I am skeptical about AI. As an athlete I would have to trust both bot’s training data and logic (much like experience and intelligence of a human coach). From business perspective, there are probably easier ways to sell more subscriptions.
Well once we get implants on our nerve system, it will record your stress levels, what you have eaten and how hard you worked out. It should then be able to predict how your legs will feel in the morning.
This very thing has been discussed elsewhere on the forum.
Beware of over-promised and under-delivered AI/Machine Learning tools.
In theory it sounds great but these algorithms are only as good as their training data. I know TR has a heap of information but whether that is useful and can lead to a definitive ‘best’ plan of training is definitely up for debate.
That said it is an exciting idea. I think they might be better served by a traditional scientific approach to dig into their data and find out what seems to work best in a given scenario.
As much as I would like training software that adjusts based of previous workouts, I think it might be impossible to create software that holds every bit of data in account.
An obvious caveat in this kind of software is the lack of ‘feel’. No software can account for how you feel and what other stressors you have in your life that may not influence the collected data. Data after all is just numbers and interpretation.
If you get experienced in training and the amount of training stress your body (and your lifestyle) can handle one can easily assess what kind of workout is feasible at what time.
Another fact will be that the guys at TR understandably will not want to be liable if something happens to a member because the software dictates some kind of intensive workout while you might be ill or not capable of doing the work.
Bottomline: learn to know assess your body (mental and fitness capabilities) and adjust training accordingly.
Don’t become completely dependant on data.
Where I’m going though, is that a real on the spot human coach from 20 years ago, had a lot less data, but could look you in the eyes and see what your mood was like. So if you instead know that for the past month, when you’re given a 20 min interval at 95% FTP, you always fall off towards the end, but this time you’re holding it … then that is a good sign that you need to be challenged more. Or even if over the past month, you do ‘‘better’’ on short intervals than long ones, then a human coach would focus more on longer intervals.
It wouldnt even need to be AI, and could just be a flowchart. Did the rider meet targets? Yes, good - keep going. No - Then had the rider done more riding leading up to the workout than planned? No. Then did the rider also fail to meet targets in the past few workouts? Yes … The riders target FTP (with their current environmental factors like sleep/diet etc) is too high, because they continually fail to meet it. Or does the rider always hit the short intervals, and always fail the longer ones? Then the rider needs the plan to focus more on adapting to longer intervals.
Without some sort of adaptation, then the TR plans are no more ‘‘valuable’’ than plans copied from a book or magazine article. The fact that they’re housed in software rather than paper, is being wasted.
I think in the 2020’s, we’ll all be approaching indoor training very differently, much like people ten years ago did it differently than we do now. I just want the 2028 experience now!!
Writing all the logic, and algorithms etc. into code might not be that easy task, let alone that all that needs to be tested so that it actually works as intended. Sure, everything would be nice to have just by a snap of fingers, but it’s not that easy.
And, companies need to also think about the ROI.
I’m sure guys in TR are doing there best to develop nice new features into the app, let’s give them some time for that.
I think the answer to the original question can be answered with another question: How much recovery do you need?
The answer always is: it depends.
For instance, if you have a Garmin smart watch, it tells me how much recovery I need after every workout I use it for. And if I listened to it, I’d be grossly out of shape.
I’m also 44 years old. I need more recovery than a 25 year old. But I also need less recovery than a typical 44 year old. I’ve gleaned this through both experience, and insight on forums like this one. Recovery is alchemy. Much more art than science.
Finally, I will say that a big glitch in the ability to use AI to adjust plans is the inherent flaws in TSS. I won’t belabor the point here, but on another thread I pointed out that I had a workout with a ~160 TSS that didn’t tax me too much, and a week later I had a race that generated a ~140 TSS, and I damn near needed a week to recover from it. It smashed me.
Once you start looking at all the variables involved with generating dynamic recommendations for each workout you start to realize how complex it really is.
I have some experience building reporting systems and some ride mapping tech. When dealing with google maps and route recommendations you start to understand that Google isn’t just giving you 3 ways to get from point A to point B. They factor in realtime traffic, weather, historical traffic patterns for the particular day, does the day of travelling land on a weekend or weekday, etc. I’m sure there’s many many more variables involved.
Trainerroad has to choose what to develop and improve. Do you shift to the “AI coach”(really just conditional clauses for a particular set of historical training variables) and risk blowing up your core app, or do you keep improving on your existing base?
Gotta love the IT buzzwords thrown around in the media as if these technologies are easily implemented. It’ll come but it’s going to take time.
@kiwifyx - really like the concept, but think there’s way more work involved than you think; take it from someone who writes software for a living, there’s a huge amount of development involved There’s also a flaw with automating the workout scheduling based solely on TR data - you only have a snapshot of data from the last workout, so it has no real idea what your physical state is when you are scheduled to do your next workout.
Maybe TR don’t actually need to re-invent the wheel here. They could instead integrate with an app such as HRV4Training - they’ve already done all the hard work in terms of accurately assessing your recovery state based on HR/HRV and health data. They also already integrate with Strava/Training Peaks etc., so I’m sure would be open to integration with TR.
Here’s how I see it working:
Every morning your recovered state (based on the HR/HRV readings and subjective health data) is automatically pushed from HRV4Training to TR.
When you load your next TR workout it takes a look at the HRV4Training data from that day and if required offers you alternative workouts. For example, TR sees that your body is poorly recovered, so instead of the VO2Max intervals workout planned, it offers 3 alternative recovery type workouts.
Here’s the bonus - subsequent HRV4Training data will be pushed to TR the following day, so it can actually see how your body has responded to the alternative workout you carried out. Over time TR learns what your body responds to best and offers you better alternative workouts.
Far less development required by TR (they simply need some code for offering alternative workouts based on the data returned from HRV4Training), but could be really effective. TR could even charge for an “advanced” level subscription for this type of functionality. I’d certainly pay a few extra £ a month for this.
If enough people used this it would also build up a really powerful TR dataset (allowing some of the ideas in @themagicspanner’s TrainerRoad’s Big Data thread to be better implemented).
I think this is a really cool idea. Even if this just started as an analysis of historical user data to see what sequence of work outs resulted in the greatest increases in performance over time. It would also be interesting to explore developing models that predicted the next best work out, or even whether the intensity of a work out was adjusted by the user. Maybe also the starting place is machine generated recommendations that are reviewed by a human, which could also serve as labels for future models.
Starting down the path of personalized training plans adjusted based on historical performance and perhaps ultimately making intra work out adjustments, including coaching advice, seems like it makes a ton of sense even if fullly automated, reliable predictions aren’t possible today.
I’d love to get my hands on an anonymized data set to build some prototype models.
There’s also a flaw with automating the workout scheduling based solely on TR data - you only have a snapshot of data from the last workout, so it has no real idea what your physical state is when you are scheduled to do your next workout.