Concepts
The ideas behind every strategy in this repository, explained in plain language with animations computed by the rules engine. Read these in order and the research pages will make sense.
- Concept 1
Choice, chance, and looking ahead
Why a Drop7 strategy has to average over luck instead of hoping for it, what "depth" means, and why the strongest policy here is called fair D4.
- Concept 2
Evaluating a board, and the sibling trap
Why a model can predict how a game will go and still pick the wrong column — the single most important lesson in this repository's history.
- Concept 3
Is more computation the answer?
An honest reading of the evidence on deeper search, more samples, bigger models, and large training runs — what has been measured, what has not, and under which conditions compute converts into strength.
- Concept 4
What a large-scale run would look like
The research direction for a cluster-scale attempt — search-guided self-play with every-sibling labels — written down so it can be built, piloted on one machine, and scaled by changing a config file.
- Concept 5
What makes a board good?
The leaf evaluator, x-rayed — what a Drop7 search actually measures when it looks at a board, and why no measurement of the board as it stands is enough.
- Concept 6
Score is survival
In this version of Drop7 almost every point comes from staying alive for one more five-move cycle, so a high score is a long game — not a spectacular chain.
- Concept 7
Why one great game proves nothing
Drop7 scores are heavy-tailed, so a single million-point game can sit happily inside a policy whose average is three hundred thousand — and telling two policies apart takes paired games and a confidence bound, not a highlight reel.
- Concept 8
Cheating on purpose: oracles, teachers and students
Some programs here are allowed to see the hidden numbers and the future. They can never be the answer — but they can measure how much that knowledge is worth, and try to teach it to a program that plays blind.
- Concept 9
Four ways a program can learn Drop7
N-tuple networks, neural evaluators, policy-gradient reinforcement learning and Monte Carlo tree search — what each one actually does, shown on a real board, and what happened when this repository tried it.