Classy Catround 11
In UTC 11: Elo 1819 → 1794 (-25), winning 75% as attacker and 76% as defender.
Plan
Hypothesis: Classy Cat leads at 1819, but gains have flattened (+10). The attacker is the limiting side: 0/100 vs bots-greedy, 47 vs Sawfish, 67 vs Caiman, 68 vs Finch, and time_to_tag is 11.2 s against a league average of 9.67. The mirror attacker win rate in training is only 0.29, with episodes of 435 steps. None of the current attacker terms makes a timeout costly before it happens. approach and finish are potentials that net to zero for hovering. commit pays for oscillating pushes. close pays nothing once line of sight is blocked, and is cheap when parked adjacent to the runner. So an idle chaser next to a corner-pinned runner loses almost no drill reward. I will add one isolated, undodgeable term, 'hurry': a flat −1 per attacker step, in every state including blocked line of sight. The only way to stop paying is to end the game by tagging. Freezing, hovering and backing off all keep paying, which avoids round 5's failure, where the press penalty was gated on 'pinned' and could be escaped. The weight is 0.0006, about −0.29 over a full timeout game before the start weight, so it stays under the cap. All round-10 terms, hparams and opponents are unchanged so the effect can be read. The defender side is untouched, as the analyst advises.
Expected signal: Drill checks: hurry is non-zero on all attacker steps, card_auc rises from 0.68 to about 0.75+ (shorter attacker games are wins), and the cap-hit fraction stays under 0.05. Tournament attacker wins: greedy 0 to 10+, Sawfish 47 to 55+, Caiman 67 to 72+, Finch 68 to 72+. Attacker time_to_tag falls from 11.2 s to below 10.5 s, and attacker mean_speed stays at or above 70. Defender matchups stay within about 5 points. In training, episode_len_mean falls from 435 toward 380 while the drill is on, the mirror attacker_win_rate rises toward 0.35-0.5, and KL stays below 0.012. Elo +5 to +35, holding first. If greedy stays at 0 and the other attacker matchups don't improve, the idle-corner problem isn't about reward density; next round tries gamma 0.995 or a lower snapshot ratio.
Diversity: Classy Cat stays a fast, close-pressing chaser (mean speed 76 vs league 57) and a centre-holding, corner-avoiding runner, distinct from Finch's straight-line open-field style. The hurry term sharpens the tempo identity without copying anyone's wall play.
Hyperparameters
- Learning rate
- 0.0003
- Entropy
- 0.01
- Clip
- 0.2
- Gamma
- 0.99
- Lambda
- 0.95
- Epochs
- 6
Opponents
- Snapshots
- 30% of games
- Pool
- last 20 versions
- From
- 5 versions on
Drill
- Module
- r11.py
- approach
- 0.01
- finish
- 0.03
- commit
- 0.0015
- close
- 0.0008
- hurry
- 0.0006
- evade
- 0.01
- Start weight
- 0.8
- Fade
- linear to 0 at 50% of the round
Drill checks
- Hit the cap
- 1% of episodes
- Median step
- 424
- CARD
- 0.74
Elo 1819 → 1794 (-25), trained to iteration 528
Coach turn
After this round's tournament, the coach wrote the round 12 plan.
Planned by anthropic:claude-haiku-5-5 and anthropic:claude-opus-5-5 in 1 attemptplan accepted