I beat Roblox's interview game with Machine Learning (adamkulik.com)

A developer reverse-engineered Roblox’s Kaiju Cats interview puzzle and built a fast simulator plus a machine‑learning search to consistently produce high scores. The game requires placing up to ~31 movement commands (6 possible actions each) to guide three cats over 15 turns through buildings, power plants (double power), traps (mud, spikes, boulders), and an order‑sensitive bed bonus (+2000, ×3, ×5). Because the real game is single‑player and slow to run, the author implemented a Java simulator (object‑oriented Tile classes, three-phase per‑turn logic: action resolution, collision handling—ties broken Blue>Red>Green—and tile effects) to run millions of trials quickly. Rather than standard RL or supervised methods, they used simulated annealing with large temperature (start ~100,000), multi‑mutation moves, and restarts, but the naïve approach plateaued around 40–50k points because single‑command tweaks couldn’t create the coordinated path changes needed to get all cats to bed. The breakthrough was a domain‑aware fitness function that penalizes distance to beds, giving the annealer a gradient to pull solutions toward coordinated multi‑step paths that trigger the huge multipliers. The project shows that for combinatorial, single‑player planning with limited real‑world rollouts, a fast simulator plus search + reward shaping (domain heuristics) can outperform blind stochastic search and be a practical alternative to expensive RL.