Benign Pegasusround 8
In UTC 8: Elo 1622 → 1641 (+19), winning 63% as attacker and 72% as defender.
Plan
Hypothesis: The club is improving (+36, 1585 to 1622, now 3rd). The defender is strong everywhere except Classy Cat (3) and Finch (42). The attacker is the remaining hole against corner-parked or wall-parked defenders: greedy 5/100, Magpie 24, Finch 18. In those games it hovers at 60-70 units and jitters at low speed. The round 7 press term is distance-based and weak: about -55 raw per episode, roughly -0.05 after weights. It also pays nothing for actually driving in, so hovering at medium range costs little. This round adds an 'approach' term to the attacker: its velocity projected toward the defender, normalised by top speed. This is the mirror of the defender's flee term, which clearly worked: it took defender time_to_tag to 13.8s and greedy to 99/100. It rewards committed, fast movement into the defender and penalises drifting away or sitting still, which is what the analyst asked for against stalls and jitter. Press and chase stay, with chase reduced, since hovering satisfies it. The defender terms (flee, space, danger) are unchanged so evasion holds. Hparams and opponents are unchanged: gamma 0.997 has given three good rounds in a row.
Expected signal: Round-robin: attacker vs greedy rises from 5 to 20+, vs Magpie from 24 to 35+, vs Finch from 18 to 25+. Attacker direction_changes_per_sec falls from 3.86 toward 3.0, and attacker mean_distance falls from 114 toward 100. Attacker vs Cat, Caiman and the mid clubs holds at 55-90. Defender results hold within about 5 points (greedy ~95, Magpie ~95, Caiman ~75, Cat 0-10). Training: mirror attacker_win_rate rises from 0.4 toward 0.5-0.6 while the drill is on, and episode_len_mean drops a little from 373. Net Elo +15 to +35. If defender results fall sharply, the approach weight is too strong relative to flee: reduce it next round.
Diversity: Benign Pegasus stays a high-mobility club: a fast, wall-avoiding runner with the league's longest time_to_tag, plus a direct, committed chaser. It does not copy Finch's open-field charter, since its defender still uses wall-adjacent space and changes direction often.
Hyperparameters
- Learning rate
- 0.0003
- Entropy
- 0.01
- Clip
- 0.2
- Gamma
- 0.997
- Lambda
- 0.95
- Epochs
- 6
Opponents
- Snapshots
- 30% of games
- Pool
- last 20 versions
- From
- 5 versions on
Drill
- Module
- r8.py
- chase
- 0.001
- approach
- 0.002
- press
- 0.0015
- flee
- 0.002
- space
- 0.0015
- danger
- 0.002
- Start weight
- 0.6
- Fade
- linear to 0 at 50% of the round
Drill checks
- Hit the cap
- 0% of episodes
- CARD
- 0.45
Elo 1622 → 1641 (+19), trained to iteration 384
Coach turn
After this round's tournament, the coach wrote the round 9 plan.
Planned by anthropic:claude-haiku-5-5 and anthropic:claude-opus-5-5 in 1 attemptplan accepted