The AI pit wall (episode 15)

From episode 15 of A Million Tries ("My AI Race Car Refused to Make a Pit Stop"). A second, small AI runs each car's race strategy, as a team's pit wall does; the race-day driver drives every car. Once a lap, 600 m before the pit entry, it decides: pit this lap or not, and which tyre; and once on the grid, the starting tyre. Evolved (mirrored-pair evolution strategies) on finishing positions: 4 of its cars against 4 run by the team's rule (pit when the worst tyre passes 60%, fit hards), random designs, random circuits, grid slots shuffled. It started as a copy of the rule. Its own idea, from generation 17: start on hards, and stop less.

128 races on 64 test circuits, final rules (8 cars a race, so 4.5 is the middle):

Read the control row: a copy of the rule already gains about half a place, because this AI decides just before the pit entry and the rule often just after it (a lap more on worn tyres). That's the design, not learning. What it learned on top is about 0.16 places. Ten more hours of training (strat-1 -> strat-2) added nothing clear.

• Sees (25 numbers): the race left, its tyres (compound, four wears, the worst, laps on the set), fuel, stops, damage, position, the gaps to the cars just ahead and behind and their tyres and stops, the circuit's length, and whether this is the grid decision.
• Does: pit or stay out, and a compound. Brain: 25 -> 32 (tanh) -> 4, plus a direct linear path (lin).
• Code at the tag episode-15: https://github.com/backstab319/a-million-tries/tree/episode-15 (agents/es_strategist.py, scripts/strat_compare.py).

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