TrainerRoad’s recent announcement of the Adaptive Training feature could mark a definitive moment in the historical progression of training for the masses. Not because they’re first - Xert, SUF, AI Endurance, VeloPro, and others have all introduced most of the features announced by TrainerRoad, and some of them have been operating for years. But as 2019-2020 saw a tipping point in virtual riding transitioning from novelty to likely a permanent and fundamental part of cycling, I think we may soon see something similar with “smart” workout and training plan generation becoming mainstream. Starting with cycling, but transitioning to other sports in time.

The subject can be daunting to approach, bringing new concepts and yet more new jargon and acronyms into training methodology - a subject already inundated with jargon, acronyms, complicated graphs, etc. But don’t despair. There may be hurdles along the way. But there should be a payoff at the end, not just in measurable performance, but in the enjoyment of the process. A guiding principle of TrainerRoad from the start has been maintaining a carefully curated, elegant user interface that doesn’t overwhelm users. From what we’ve seen so far, that principle is being maintained with Adaptive Training. And the other “smart” training platforms have similar elegant interfaces.

This article is intended to be a gentle introduction into what intelligent workout and training plan generation is, why it might improve some aspects of training, how we can prepare ourselves for these systems, and how we can evaluate them. It is not a deep-dive into the details of daily use of TrainerRoad’s Adaptive Training - that will come later after there’s been some time to collect experience with it.

As a note for those curious to try Adaptive Training, TrainerRoad says that the beta testing program will not be a long-running process. It is only being used to hammer out any lingering bugs, and full Adaptive Training will be released out to the masses relatively quickly.

What is Adaptive Training?

TrainerRoad is rolling many features into the brand name “Adaptive Training.” I recommend listening to episode 298 of the Ask a Cycling Coach podcast for a long, detailed introduction to most aspects of it, present and planned. DCRainmaker also has a nice introduction to the TrainNow feature, the one feature available for all TrainerRoad users without the need for special access to the Adaptive Training beta testing program (as of the writing of this article).

But what problem are all these features attempting to solve? One answer to that is that it’s an attempt to solve a long-running criticism of TrainerRoad: workouts and plans that are often just too difficult to complete as prescribed. There are myriad reasons that an athlete could have difficulty with TrainerRoad’s “old” system. Just a few examples are selecting a workout or plan not consistent with their current fitness level, being suddenly inundated with life stress or events, being a really strong performer on Ramp Tests resulting in over-estimated FTP, or having different responses to training stress compared to the “average” person. Adaptive Training claims to be able to gracefully handle all of these situations, and more. The goal is for athletes to be far more likely to have the satisfaction of workout and plan completion, while still being challenged sufficiently to achieve progression and reach goals (more on progression below).

Another, more philosophical, answer is that it’s much more than that. It’s not just an attempt to improve TrainerRoad’s rate of workout and plan completion, but a sea change in training methodology. In the quest for improved fitness and competitive success, most athletes use a combination of analytics tools, personal experience, rules-of-thumb, and coaching. We often copy patterns that other people seem to use successfully. There is plenty of “science” to help guide us, but scientific studies usually answer narrow questions. They don’t tell us how to live daily life as athletes. But successful athletes are the ones who can adapt all of these sources of information into a consistent set of daily practices, responding to each new piece of information intelligently. Managing this is often more like art than science. And because there’s so much human judgment involved, we are subject to the human failures of bias, ignorance, and even delusion (Despite over 30 years as an endurance athlete and coach I still frequently overestimate my capacity to handle training stress and race frequency).

Machine learning - a form of artificial intelligence (AI) - can conceivably handle some of these problems better than we can. Maybe better than the best coaches can. While the human mind is good at some kinds of problem solving, there are some things that computers can simply do better. Machines are good at finding subtle but persistent patterns in large volumes of data that we can miss. Just as one example, the “ramp rate” - or the rate of increase of training stress - is a tool that some data-minded athletes and coaches use when creating cycling training plans. And there are all manner of rules of thumb and expert knowledge about what the “right” rate should be for a given athlete on a given day or week. The rate is either applied as a static rule of thumb over an entire workout plan, or adjusted as the plan progresses by looking at past workout performances and often direct conversation about how the athlete is feeling. A machine learning system may one day be able to tell us with what ramp rate an athlete can handle - or even better - than most humans can. It may be able to find patterns in data that - after having learned from processing data from thousands of athletes - reliably indicate when an athlete is approaching burnout, even before the athlete feels the tell-tale signs of overtraining. Or, conversely, invisible (to humans) patterns that indicate the athlete is ready for a breakthrough.

And I think this is what TrainerRoad is attempting here: the changes approach the equivalent of having a coach take a look at an athlete’s “analytics” before each workout, and then review workout results and have a quick chat after each workout. Then the coach would use all that information to make suggestions about how to approach the next workout. That consistent daily feedback is something that some data-minded athletes with free time and familiarity with cycling analytics can do themselves, but is otherwise the domain of the elite or wealthy athletes with the luxury of daily coaching. If Adaptive Training works well it could be an tremendously beneficial tool for the mainstream athlete.

What Changes in Day-to-Day Use?

Regular users of TrainerRoad should be able to seamlessly transition into use of Adaptive Training. There is no huge change to the usual patterns of interaction with the calendar, career page, plan generation, or workout selection. One difference is the addition of a “Progression Level” chart.

These levels show how an athlete is advancing in each of one of the seven TrainerRoad effort levels. And one minor criticism I have is that that seven numbers is...a lot of numbers. Is it useful for an athlete to know how they’re progressing in “sweet spot,” which is presumably highly correlated to threshold? Though I understand this is likely just TrainerRoad’s adherence to their historical use of seven levels for the sake of consistency. Another slightly confusing aspect is that these numbers are not absolute, nor description of fitness - e.g. having a “9” in threshold indicates nothing about threshold power. It just shows progression in ability to handle threshold-heavy workouts. And the numbers will change any time FTP changes, e.g. after a Ramp Test. Though it’s a big TrainerRoad point of emphasis about Adaptive Training that large changes in FTP from a Ramp Test no longer result in big changes in workout difficulty. The progress should always be manageable, without large discontinuity.

So the numbers are snapshot in time, and useful only for personal workout selection and feedback about recent progress through a training plan. That makes the chart completely different from the likes of SUF 4DP which is an absolute description of performance in four power durations.

Every workout is now scored by the Progression Level. Here we see that this workout has a Sweet Spot level “4.7.” . This “4.7” is intended to be compared to the value in the corresponding “sweet spot” Progression Level. An athlete can “grow” their Progression Level by completing workouts that challenge that effort level - the higher the value attempted, the more potential progression there is. It’s a bit of “gamification” of workouts, though with far more substance than the “gold stars” received Zwift for successfully knocking out intervals.

For ease of interpretation, the workouts are also assigned a qualitative assessment of difficulty according to a new taxonomy (from easy to hard): Recovery, Achievable, Productive, Stretch, Breakthrough, and Not Recommended. This new vocabulary provides a quick assessment how the workout should “feel.” I am eager to see how well these new keywords match my subjective experience. It is encouraging, though that TrainerRoad lets users veer “off the reservation” and try brutal workouts. Most coaches and athletes have at times experienced the value of the “breakthrough workout” in developing confidence and helping trigger escaping plateaus in performance.

Lastly, the feedback that Adaptive Training requests is also encouraging. This comes in the form of giving the athlete the power of either accepting or rejection suggested changes in “levels” and giving short surveys on workout performance after each TrainerRoad-generated workout.

The purpose of this feedback is probably two-fold. The first is that it’s a recognition that machine learning may at times get it wrong. Sometimes we know better when we’re ready to take on more stress. Or at least we’d like to accept that additional stress on our own terms. Another is that it likely serves as feedback to TrainerRoad’s algorithms, allowing the system to tune itself to the patterns of performance an athlete may have.

Are FTP, TSS, etc. Dead?

Absolutely not. FTP is still the primary performance measure within Adaptive Training. There are still Ramp Tests to estimate FTP. Though Adaptive Training introduces machine-learned FTP estimates and FTP prediction as well. The rest of the classic Banister impulse-response model analytics, and Dr. Coggan-derived Performance Management Chart (PMC) and other tools are still there. And though TrainerRoad is very secretive about what machine learning “features” (in machine learning techno-jargon) they use, I’d bet some important ones are based on these classics of cycling training analytics. The Banister model, TSS, CTL, are probably integral to the machine learning process. TrainerRoad is probably building on training science more than scrapping it and starting something new. The hope may be that “AI” can manage and interpret a PMC better than we can.

Is it Smarter Than a Coach?

Even if these systems work as well as advertised, most human coaching is not under threat, even over the long term. At least human coaching beyond just doling out weekly template-based workout schedules, which I’d barely call “coaching” to begin with. Technology usually doesn’t displace human labor - it frees humans to spend more time on the things we’re better suited to perform. Today’s farmers can spend less time behind a plow horse, more time doing analytics on which crops will be the most profitable to grow. These systems are likely to become another tool for coaches, providing more insight into athletes and how they’re doing. There are still myriad decisions to be handled that are not fully answered by “AI” and are unlikely to be anytime soon: how to arrange training macro cycles, how to schedule races, when to take significant downtime, bike fit, aerodynamics, swim technique, race strategy, unravelling anxieties or unlocking wellsprings of motivation, etc. I could go on for pages.

It took decades of development for “AI” to consistently beat the best human chess players, and as complex as chess is, it’s arguably simpler than managing human beings and human physiology. Athletes will still benefit from expert guidance for the foreseeable future.. Without going into eye-glazing techno-jargon, there are myriad ways that machine learning can perform poorly or even spectacularly fail, and Adaptive Training will not be immune from these failure conditions, and will likely never completely replace our own human insight as athletes and coaches. Coaches who understand how to use these new tools for their own benefit, however, might gain competitive advantage.

What's Missing?

Adaptive Training is for cyclists. Don’t count on multi-sport anytime soon. TrainerRoad, in the aforementioned podcast, did provide a mea-culpa in perhaps “leading on” triathletes a bit too much. There’s no reason a triathlete couldn’t use the platform for cycling training, though.

There is no macro-cycle planning, to my knowledge. Meaning that the system indicate when to take an easy week. It should make recovery workouts appropriately manageable, but that’s it. And there are other very important aspects of plan creation and execution that are subject to solely “human knowledge.” At least for now.

As of yet, there are no feedback surveys for outside workouts, e.g. outdoors rides, Zwift, etc. These workouts are incorporated into the machine learning process, and they do influence level progressions. They’re just not yet fully incorporated - so for the moment TrainerRoad tends to prefer its own ecosystem to take fullest advantage of Adaptive Training.

To touch on a continuing TrainerRoad controversy, Adaptive Training will not figure out whether “sweet spot” vs. “polarized’ training is better for a given athlete. Or any other pattern of prescribing workout stress. It’s still up to the athlete or coach to navigate many aspects of plan generation. However, TrainerRoad is starting a new polarized training option as a way of exploring its use, and determining its efficacy vs. sweet spot among TrainerRoad users.

How Do We Know It Works?

This question will invariably lead to controversial debate for very good reasons. Buckle up and come to the forum thread for a spirited discussion!

TrainerRoad’s system is unlikely anytime soon to be validated in the same way that the work of say, Dr. Banister and Dr. Coggan were. Or in the way that most broadly accepted training science is vetted. The simple reason is that TrainerRoad’s system is almost entirely proprietary. And their database of user data is private and closely held. The internal workings are unavailable for inspection or critique. There are perfectly good reasons for this. TrainerRoad spent a huge amount of their own money and time developing the system and database. Successfully protecting their intellectual property may be necessary to survive competition from the likes of Zwift, not to mention all the other players in the AI training market. And TrainerRoad is not alone here. To my knowledge, few of the players in the “AI training” are making their work available for peer review. This is unfortunate for the advancement of training science in the near term. We can only hope that at some point once these companies establish secure positions in the market they can begin to publish pieces of their work for review. But until scientific peer review is possible, TrainerRoad - and the others - will be subject to criticism and speculation about transparency and the justification for marketing claims.

So absent broad scientific validation, we are left with trust, anecdotal consensus, and perhaps isolated testing of some of the predictive aspects of Adaptive Training. TrainerRoad has significant credibility in the “trust” department. They’ve earned a reputation over the years for being methodical, reasonably transparent, consistent, and for generally following through on major claims. They also, in my opinion, have the science and engineering “chops” to attempt ambitious projects like this one.

We will certainly get anecdotal information about how well the system might work. With TrainerRoad having a loyal following in general, one anecdote will not have much worth, but if over the next year or so we see lots of people give positive feedback - particularly those with no prior TrainerRoad experience - it will be hard to ignore.

And for those who prefer quantitative measurement, we may be able to independently test some of the predictive aspects of Adaptive Training. Adaptive Training will predict FTP within a probabilistic range of some kind. (I do not yet have beta access, so don’t yet know all the details). We should be able to independently validate those predictions vs. an FTP test or time trial result. I look forward to putting some of these claims to the test. This is not true scientific validation, but if the claims match reality, it could significantly increase confidence in the validity of the Adaptive Training system.

If Adaptive Training is smoke-and-mirrors (which I doubt) - the truth will be outed in time. Helping thousands of athletes improve their training, race results, and enjoyment of the process is not something fakeable over the long term.

How to Prepare for AI Training

Even someone who doesn’t subscribe to any of the AI training ecosystems, or doesn’t yet have access to TrainerRoad’s beta program, can start preparing for better use of any machine learning-based training system: embrace the use of sensors. This includes sensors on bikes, wearable sensors, phone apps, day, night, indoor, outdoor. Cycling power measurement is by far the most important for the present. But all those power meter zealots who ditched their heart rate straps 10 years ago, might want to consider bringing them back. Or buying new ones capable of HRV and/or ECG measurement. Sleep trackers may also end up being useful. Other possibilities include daily weight measurement, nutrition and hydration tracking, menstrual cycle tracking, blood sugar tracking (particularly for diabetics), etc. The possibilities for wearable tech and “quantified living” are expanding rapidly.

Even if this data isn’t used right now (and few of the sensors mentioned are used yet), having a long baseline of recorded data may eventually pay off. Not all of these may end up being very useful for the purposes of optimizing training. But it’s difficult to predict which ones will or won’t.

Embracing “quantified living” should fit into personal comfort levels for privacy, cost, and hassle. It’s important that the resulting data be recorded digitally, and is usable by third parties. It can be useful to use one of the larger wearable tech ecosystems like Garmin Connect, Google Fit, or Apple Health. Whichever method used, be very careful to make sure that YOU own the data. This means that at any time you can download all your data for your own purposes. Otherwise it may never be usable by a third party such as TrainerRoad. And TrainerRoad eats their own dog food in this regard - they allow users to download all their raw data upon request.

The Future

I have read some forum comments already to the effect of, “I’ve been training for 30 years, and I know myself so well that there’s no way AI could ever be of benefit to me.” Those commenters may be right to some degree. Machine learning may never replace those 30 years of experience. But I’d warn against discounting the system entirely. When machine learning is expertly applied to a problem that it’s well suited for, its performance can be absolutely astonishing, and exceed human performance in some ways. I’ve experienced this astonishment while testing prototype self-driving cars. I have yet to experience this astonishment with TrainerRoad Adaptive Training or any other “smart” training platform: I’ve found some of them useful and interesting, and am still waiting for the experience that cements their value for me. It’s possible they’ll fail because human training just isn’t well suited to the tools of machine learning. But I’m leaving myself open to their potential. I think there’s a strong chance these systems will grow and significantly improve our training, race performances, and lives. I look forward to testing and watching their progress. Count me as an early adopter.