Inspiration
Build a Tron-style agent that “reads the crime scene” fast: stay safe, claim space, and trap opponents—under strict CPU/Docker limits.
What it does
- Plays the 18×20 arena with a strict safety filter
- Scores moves by space/Voronoi, branching, tunnels, and choke points
- Uses tiny beam look-ahead and a smart BOOST policy
- Hot-tunable weights in
weights.json+ auto-training loop
How we built it
Python + Flask agent, state.py for wrap/legality, heuristics.py for scoring/search, and parallel self-play tools to mutate and promote better weights.
Challenges we ran into
Mirror crashes, wrap (x,y↔r,c) bugs, slow judge loops, and determinism—we fixed with stricter safety, coord normalization, a fast offline trainer, and light exploration noise.
Accomplishments that we're proud of
Reliable safety gate, territory-aware planning that wins endgames, and a 24-thread autoloop that steadily improves weights.
What we learned
Safety pruning + small look-ahead beats heavy ML here; space control > short-term greed; hot-reloadable weights speed iteration.
What's next for Case Closed Challenge Bot
Deeper opponent modeling (mini-max/MCTS within time), scripted sparring bots, endgame fill mode, and quick dashboards for weight sweeps.
Log in or sign up for Devpost to join the conversation.