Open loop

Decide a whole short sequence of columns in advance, judge it against many imagined disc streams at once, and play only its first move before planning again.

The intuition

A look-ahead search branches: if the next disc is a 3 I play here, if it is a 6 I play there. That is the right model of the game, and it is expensive — the tree multiplies by seven at every chance point.

An open-loop planner refuses to branch. It writes down one plan — a fixed list of columns — and then asks how that single plan performs across a whole set of imagined disc streams. Because one plan must work for all of them, it cannot quietly assume it will know something it will not know. The cost is that it also cannot represent the value of being able to react, which is real: a good player's plan genuinely does change when a 1 arrives instead of a 7.

Since the planner replans from scratch every move and only ever plays the first column of its plan, the reaction happens anyway — one move late.

How it works, step by step

1. Read the public position — board, next disc, rise clock.
2. Draw a set of scenarios: complete imagined streams of future discs and gray-disc reveals, generated from a solver-local seed. The same scenarios are used for every candidate plan, so plans are compared on identical luck (common random numbers).
3. Grow plans one column at a time. Extend every surviving plan prefix by each legal column, simulate all of them through every scenario, keep the best handful, and repeat to the plan depth. That shortlist is the beam.
4. Score a plan by the average value of the boards it reaches, optionally penalised by the spread across scenarios, so that a plan which is good on average but erratic loses to a steadier one.
5. Play the first column of the best plan and throw the rest away.

The leaf value can be the ordinary hand evaluator, the recursive-potential evaluator, or the tunneling residual; the lab compares all three.

What happened, in plain English

There is no result here, and the index does not claim one. This lab is classified support-only: it exists so that other policies have something to be measured against, and its most visible role in the record is as the comparator that tunneling improved on in a 64-game comparison whose numbers were never retained.

It is also one of the six policies you can watch play in the benchmark playground, which is the easiest way to get a feel for how a plan-then-replan policy behaves.

The technical record

The experiment index records this lab as support-only, repository-verified: "it primarily serves as a comparator." Repository-verified means the implementation is present and its purpose is established from the source; it does not mean a run happened. No cohort, score, or comparison for this lab is recorded in the experiment history or anywhere else in this repository, so this page reports none.

Repository-verified from the source. src/core/typescript/robust-open-loop-beam.ts takes a scenario count, a plan depth, a beam width, an optional risk aversion ("mean − riskAversion × population stddev"), a deterministic work cap and a loose wall-clock cap, and hard limits of 8 plan columns, 256 scenarios, 256 beam entries, 10,000,000 work units and 60 seconds. The lab main.ts defaults to 16 scenarios, plan depth 4, beam width 32, risk aversion 0, a 2,000,000 work cap and 4 games from the 0x1d70… training range, comparing the combined, recursive and tunneling leaf profiles. The registered playground policy open-loop-beam uses 12 scenarios, depth 3 and beam width 8 under a 150,000 work bound.

The planner's scenarios are generated from a solver-local seed and the observable state. It never receives the game seed, the hidden gray values, or the real future disc tape — the scenarios are the planner's own guesses, and they are wrong in exactly the way a player's guesses are wrong.

What this taught us, and what is still open

• Support code is not a result, and labelling it honestly matters. A comparator that has never been measured against the reference search cannot certify anything it wins against. That is why the tunneling comparison is described here as a comparison and not as evidence.
• Open loop is a different failure mode from determinization. Imagining a complete future and planning perfectly inside it produces plans no real player could have made — the failure the MCTS work calls strategy fusion. Scoring one fixed plan across many futures avoids that by construction; what it gives up instead is the value of reacting. Neither effect has been measured here.
• What is still open is whether a plan-shaped policy is worth anything at Drop7 at all. On the evidence in this repository, that question has not been asked.