Skip to content
All essays

Tennis3 min read

Tennis training through the lens of a neural network

用神经网络的视角理解网球训练

Treat every swing as a parameter update and every match as a test set. It did not make me a better player, but it made practice much less likely to wreck my mood.

Abstract cover: a fine line grid and concentric circles

For a while, every session ended the same way. In feeding drills I looked fine. In a match I fell apart: the movement broke down, the ball went everywhere.

I assumed it was mental. Later I decided it was not my head. It was how I practised.

Feeding drills are great; matches fall apart

A coach feeds from the other side. The placement is fixed, the rhythm is fixed, the spin is fixed. I hit eight out of ten.

A match is a different thing. Some balls are deep, some are short. Some spin, some do not. The rhythm keeps changing. I was still playing the movement I had trained, so I made errors.

In machine learning this is called overfitting: good on the training set, bad on the test set. Feeding drills are the training set. Matches are the test set.

The fix is not more feeding. It is variety. First fix the placement and get the movement right; then slowly add different spin, depth and pace. My coach already did this. He just did not have the word for it.

The clearest example was a university match in second year. In practice I could rally for twenty shots. In the match I was broken in the first service game. After four games I sat by the court thinking: what exactly have I been practising?

Movement is a parameter

Picture a person as a function: it sees a ball, it produces a movement. The mapping in between is your parameters.

Practice is parameter tuning. A coach correcting your movement is giving you a signal that something is wrong. Repetition takes the right movement a few more steps. The day it suddenly clicks, the parameters have landed somewhere useful.

The useful part of this framing is that it turns “I played badly today” into “I handled this ball badly.” The first is an attack on myself. The second is a search for the problem.

It sounds cold, but it cut down a lot of self-flagellation. Before, one mishit could keep me awake. Now I ask: was it a bad read, or did the movement not follow?

The most useless sentence in tennis is “I didn’t have it today.” It explains everything and says nothing.

One variable at a time

I used to be greedy in practice. Fix the forehand today, the serve tomorrow, the net game the day after. I tried to fix everything at once and half-fixed all of it.

Now I have one rule: one variable per session. This week only the contact point. Next week only the weight transfer. Everything else can look ugly. It feels slow, but at the end of every session I know what I changed.

Learning rate works the same way. Change too much at once and the parameters swing back and forth and nothing converges.

How to use it in a match

Thinking about parameters is fine in practice. In a match there is no time.

In matches I do one thing: between points, decide how to play the next one. Where to serve, which side to attack, whether the opponent’s backhand or forehand is weaker. One thought. More than that and I get confused.

After losing a few points I stand behind the baseline and breathe twice. That trick came from my notes — I noticed that when I lost several points in a row, I was usually still thinking about the last mistake.

Tracking three things

If you do not write it down, it did not happen. After each session I note three things:

  • Result: how many games won, how many lost, how many unforced errors.
  • Movement: how I would score the one thing I practised today, one to five.
  • Feel: was it light or forced, was I attacking or just surviving.

Keeping them separate stops one lost match from invalidating the whole practice block. Sometimes the score is bad and the movement is right. Sometimes you win because the opponent gifted you points.

It does not need to be formal. A few lines on my phone take five minutes. A month later I can see which phase I was stuck in.

Last month I noticed that I make far more errors when the opponent’s return is deep. Before, I would have called it “bad hands today.” Now I know which ball causes it, so practice has a target.

The side effect

There is a cost to seeing everything as tuning.

Once, in a match, I stood at the baseline thinking: which set of parameters should I use here? The ball arrived while I was still thinking, and I got passed.

I learned that you can tune in practice, not in a match. In a match, one or two thoughts: watch the ball, weight. Any more and the movement slows down.

So theory is for off-court. On court, simpler is better.

Not just tennis

Later I noticed the same approach works elsewhere.

When code gets stuck: is the input wrong, or is my processing wrong? When a customer conversation goes badly: was the script wrong, or did I not understand the actual problem?

Breaking “I’m bad at this” into “which step is failing” is probably the most useful habit I picked up in these years. It does not guarantee a solution, but it tells you where to look. I now ask “which step” whenever something blocks me — code, customers, writing.

This framework has not made me a good player. I still get run around the court.

But it changed one thing: after losing, I no longer wonder whether I am built for this. I wonder which parameter was off and what to change next time.

One consumes emotion. The other produces movement. That might be the most useful thing tennis has taught me.

The framework is not magic. Some balls are just bad. Accepting that is part of training too. My goal now is simple: play twice a week, do not get injured, improve slowly.