Sharpen a soccer striker with evidence — then run a tournament.
Today's superpower
Rematches are science
The Arena is deterministic: same start → same result, every time. So when you change your brain and rematch, any improvement is real, not luck.
Train → save → rematch → keep what helped. That's the loop.
replay it — identical
Which recipe wins? (one run, 72 games each)
Imitate an expert first 🏆
Recipe
Win-rate
Teach (copy an expert)
20% — strong, fast
Evolve from scratch
0% — much weaker
Copy a good dribbler instead of discovering from nothing. Train well to win soccer: hand-coded strikers lose outright; the best recipe here was Teach → Evolve vs the opponent (47%).
⚠️ More training ≠ better
Refining can make it worse
Take a strong distilled striker, then…
Win-rate
nothing (Teach only)
20%
Q-Learn (reward-refine)
0% / 4% 👎
Both Q-Learn variants (vs a cone / vs the live opponent) landed below Teach. A solid imitation policy is fragile — a second objective can pull it off its good habits. Keep the version that wins the rematch, not the one that "trained the most."
🧪 One match isn't an eval
Why we don't trust a single game
An early one-match test on a device told a tidy story about a recipe. Over 72 matches it didn't hold up.
The Arena is deterministic — so one match is reproducible. But reproducible ≠ representative. Always average over many seeds.
You can watch it happen
Over-fitting 🥶
Evolve too long on one fixed pitch and the bot “just stays there” in a real match — it memorised that board.
Fix: vary the board while training. (Sound familiar? It's the Day-2 maze lesson again — a memoriser fails on the unseen.)
the population evolving
The loop that levels you up
Lost? Train against the winner
Save the foe — it copies the winner into your library.
Train against that exact bot.
Rematch (it's deterministic — a fair test).
Caveat: a big gap can need more than one pass.
train · save · rematch
🩺 “My bot got worse!”
That's the best lesson, not a fail
Refining often doesn't help → keep your Teach base; only adopt a refined brain if a rematch proves it's better.
Evolved long on one board? → vary the board (over-fitting).
Judge a change over many games, not one.
Still stuck? → re-Teach a fresh base — more training isn't the fix.
The capstone
Tournament 🏆
One device: vs Computer → Tournament (Cup or Ladder).
The class: Arena → Radio → Room — every board runs the same bracket and crowns the same champion.
No internet, no scoreboard — nothing leaves the room.
Arena → Radio → Room
The whole week in one question 💭
“Hand-coded bot or trained bot — which do you bring, and why?”
Game
Who wins
🧩 Maze
✍️ Hand-coding — a correct rule generalises
🤖 Battle
✍️ Hand-coding — clear priorities win
⚽ Soccer
🧠 Learning — if trained well (hand-coded bots lose 0%; trained wins outright)
Match the method to the problem.
🏁 You did it — all five days
Champion 🏆
I trained a soccer striker myself.
I sharpened it with a recipe I can explain.
I know when to write rules and when to train them.
From “I write the rules” → “I train the rules.” You built and trained a robot. 🤖⚽
