Game Lab · Learning your habits

Goalie vs you

I'm the goalie! I have to dive before you kick, so I can't see the ball. My only trick: learn your habits. Tap the goal to shoot!

Goalie's guess: where you'll kick next Saved Goal

Saves per 10 kicks (dashed grey = lucky guessing, 17%)

Checkpoints — the goalie's brain saved along the way. Tap one to see what it guessed back then.

How it works

Watch

The goalie remembers every kick: where it went, what came before, and whether you scored.

Look for clues

It has five little detectives. Each one has its own clue, like "what does this kicker do after scoring?"

Trust the good detectives

Detectives that guessed right get trusted more. Wrong ones get trusted less.

Dive

It dives where the trusted detectives point. Old memories slowly fade, so new habits get noticed.

What just happened?

For grown-ups: what's really going on

This is online sequence prediction, the same idea as predicting the next word, but for penalty kicks.

  • Experts: five frequency models over the 6 zones with different contexts: none (marginal), the last kick (order-1 Markov), the last two kicks (order-2), the goalie's last dive, and the last kick plus whether it scored. Counts use add-0.3 smoothing.
  • Forgetting: each context's counts decay by 0.93 per observation, so the model tracks non-stationary behaviour (a changed habit).
  • Mixing: a multiplicative-weights / Hedge mixture: after each kick, expert e's weight is multiplied by (6·pe(zone))0.6, renormalised with a floor so dormant experts can return.
  • Decision: dive to the argmax of the mixed distribution. Chance is 1/6 ≈ 17%.
  • Curving ball: shots land one zone right of the aim with probability 0.6. The goalie only sees where the ball lands, so the best possible save rate is capped below 100%.
  • Against a truly uniform-random kicker no predictor can beat 1/6 in expectation, which is why mixed strategies are optimal in real penalty shoot-outs (game theory).