Repository navigation
Lookahead: the model gets the drones' flight (-0.47 to +0.05) - #136
Merged
Merged
Conversation
…ndings The model counted a drone in a tower's circle as soon as it was near; in the arena it comes down from its orbit height first. The lookahead went from -0.47 to +0.05 against a commander with no advisor. Drift, redeploy and charging are switches of the model; the example measures each. Co-Authored-By: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Why the lookahead lost
The lookahead model had no flight. In the model a drone counted in a tower's circle as soon as it was within 60 m. In the arena it flies at its orbit height (55-87 m) and comes down at 3 m/s only within 60 m of the tower, and only a drone below the treetops counts. That is 12-22 s of nothing at every arrival, which the model did not know. Moving a drone that sits low in a circle to another tower was always a gain in the model and a loss in the game.
Measured on the old model (40 seeds, both sides; each taken proposal replayed from the seed with and without it, nothing else proposed for 45 s): 45 proposals made and 25.5 taken per game; promised +17.6, realised -1.9. Drones that were capturing: promised +10.9, realised -16.5 ± 1.3 (204 moves).
What changed
lookahead.rs: the model has height, climb, descent, take-off and charging drones (aModelof switches;driftandredeployare there but off, because they measured worse).Lookahead::with_model,with_cooldown,Model::spare_capturing: for measuring; the defaults do not use them.cognition_towers --example lookahead(calibrate,variants).18-advice.expect: the stale-proposal expectation now says west tower (the old lookahead used to move the drone to the centre there) and the text says what the model keeps.plans/actors.mdstep 4.Numbers (400 seeds, both sides, vs no advisor)
Oracle enemy positions: +0.05; blind: -0.03 (belief is not the problem). Delays now: 2 s -0.16, 5 s -0.27, 10 s -0.55, 30 s -0.94.
judge --game captain --champion timid --challenger default --seeds 24: lookahead as it was +0.175 (lower bound -0.132, kept out); with the fix +0.384 (+0.048, promoted);SIM001_LOOKAHEAD=0+0.268 (+0.031).What did not change
The lookahead is level with no advisor on this arena, not a gain. I did not tune the margin (5 points would give +0.18) and did not touch the default; whether the default commander keeps the lookahead is the owner's call. Still unexplained: why moves of capturing drones still lose (-6.4 ± 1.5), and why drift and redeploy measure worse.
Tests:
cargo test --release -p recreation_towers -p cognition_towerspasses; clippy clean for the repo's code.🤖 Generated with Claude Code