Sixteen years ago, I wrote a Java program that tried to pick the ideal move in a puzzle game called Drop7. On each turn, the player has only seven choices for their next move, but an element of chance turns this simple game into a surprisingly difficult problem involving search, hidden information, long-term planning, and chance.
I renewed my investigation this weekend, asking whether there is an autonomous strategy that averages more than one million points per game. An experienced human player can reach a million points with long-term strategic thinking, but the best automated strategies I could come up with are currently averaging around 300k-400k points.
The strongest research candidate so far, learned lookup tables inside a depth-4 search, averaged 516,155 points over 512 unseen games. A depth-4 expectimax with a hand-written leaf averaged 382,567 on the same games. Both remain well under a million, so the research problem is still open.
One of the easiest ways to contribute is to compete on the global leaderboard and submit your human strategy at drop7.dev/compete. The project has also been an experiment in using agents as research collaborators and evidence collectors. The full repository is available on GitHub under keshavsaharia/drop7-bench.
Feedback is especially welcome from people knowledgeable in game theory, stochastic planning, learned value functions, or heavy-tailed evaluation. How would you approach long-horizon action ranking here?
Source: Hacker News · Summarized by HeadlinesBriefing