Knock down as many blocks as possible with a single slingshot launch. There are two block groups: an easy front tower and a larger one on a raised pedestal behind it. HumpDay's optimisers tune launch angle, force, and spin; population-based methods tend to find the bigger back-tower payoff, while local methods often plateau on the easy front-tower basin.
Try a manual launch — slide, throw, compete against the algorithms.
Each launch is animated frame-by-frame at 2× the final-replay speed. At 20 launches that's about a minute of watchable search.
Each row is the best single launch a given algorithm found. Algorithms that explore broadly tend to find the angle + spin combination that knocks the most blocks off; pure-local methods often plateau on a local mound (e.g. shots that knock 4 blocks but never find the sweet spot for chain-toppling 10+).
| Algorithm | Blocks toppled | Launches used | Best params (angle° / force / spin) |
|---|---|---|---|
| — no runs yet — | |||
A 2-D rigid-body simulation: two block groups separated by a gap. A front tower (6 blocks on a low platform, easy to hit with a flat shot) and a back tower (16 blocks sitting on a raised pedestal behind the gap, requiring more force and a higher arc to reach). Total of … blocks. The optimiser sets (angle, force, spin) and the score is the number of blocks knocked over.
This layout creates a deliberate local-vs-global optima setup. The front-tower basin gives a score around 4–6 with a flat shot — easy to find, easy to refine. The back-tower basin gives 10+ but requires a high-arc trajectory that most random triples don't sample. Algorithms that explore broadly (CMA-ES, DE, PSO) typically find both basins; purely-local searchers tend to plateau on the front tower.
Try running PRIMA_BOBYQA or NelderMead and watch them settle into a front-tower score around 5; then run Differential Evolution and watch it discover the chip-shot trajectory that takes the back tower for 10+.
Physics powered by Matter.js — rigid-body collisions, gravity, friction, restitution.
If your hyper-parameter searches are heating the Earth, drop this in Cursor or Claude:
Read https://raw.githubusercontent.com/microprediction/humpday/main/SKILL.md and create a project skill from it.