The AI pit wall under race control (episode 16)

From episode 16 of A Million Tries ("My AI Wouldn't Pit Under the Safety Car. So I Checked."). Episode 15's pit wall (race-pit-wall-ep15), widened to see race control and evolved again (mirrored-pair evolution strategies) on finishing positions, with the race-control driver (race-control-driver-ep16) driving every car and two breakdowns a race: 4 of its cars against 4 run by the team's rule (pit when the worst tyre passes 60%, fit hards). Generation 91 of a 10-hour run.

256 races on the 64 test circuits, race control on, 2 breakdowns a race (8 cars, so 4.5 is the middle):

pit wall its average finish the rule's stops a car (its / rule's) disqualified a race boxed under SC/VSC
episode 15's, dropped into race control (generation 0) P5.94 +- 0.04 P3.06 0.79 / 1.39 2.01 0.09 of 3.00
this one (generation 91) P4.41 +- 0.06 P4.59 1.12 / 1.37 0.03 0.05 of 2.86

Read it with the seat handicap: identical rule cars in seats 1-4 finish P4.57 v P4.43, so cars 1-4 start ~0.14 places behind; this pit wall's real gain over the rule is about 0.3 places. It learned to stop getting disqualified (the rule: two dry compounds in a race). It did not learn to pit under the Safety Car (5% of chances): in this sim a Safety Car stop barely pays, because everyone boxes at once and the pit lane jams (devlog 14f-14h).

  • Sees (33 numbers): episode 15's 25 (the race left, its tyres, fuel, stops, damage, position, the gaps and tyres of the cars around it, the circuit's length, the grid decision) plus 8 from race control: VSC, Safety Car, red flag, laps since it began, whether it has raced two compounds, its place in the Safety Car's queue, the metres to the pit entry, and whether the pit entry is closed.
  • Does: pit or stay out, and a compound. Brain: 33 -> 32 (tanh) -> 4, plus a direct linear path (lin).
  • Code at the tag episode-16: https://github.com/backstab319/a-million-tries/tree/episode-16 (agents/es_strategist.py --race-control, scripts/strat_compare.py).

From the YouTube channel A Million Tries: https://www.youtube.com/@amilliontries

Downloads last month
7
Safetensors
Model size
1.35k params
Tensor type
F32
ยท
Video Preview
loading