Section 8.5, paragraph 3Whether AlphaZero-style self-play discovers the constructions is open
Open or deferred item Explicitly left open, deferred or not yet connected.
1 Paper text
The immediate experimental question is whether AlphaZero-style self-play can discover the restoration rule, complementary rollout policies or recurrent accumulator without their supervised targets. A follow-up should compare outcome-only training with curricula and auxiliary supervision, and examine the learned computations as well as win rates. For recurrence, this includes accurate accumulation and reliable state retention across separators. These controlled NIM tests would establish a basis for the harder benchmarks in Section 8.3; efficient self-play learning and the training regimes needed to achieve it remain open.
Generated from paper_jmlr.tex, lines 1261–1261.
2 Explain this
The open experimental question: can AlphaZero-style self-play discover the restoration rule, complementary rollout policies or the recurrent accumulator without supervised targets? A follow-up would compare outcome-only training with curricula and auxiliary supervision and inspect the learned computations, not just win rates. For recurrence it would also test accurate accumulation and state retention across separators.
3 How this is formalised
Status: open (stated as such in the paper). No theorem or experiment here addresses it.
| Paper anchor | sec:opt-gap-p3 (website anchor; the manuscript has no label here) |
|---|---|
| Displayed as | Section 8.5, paragraph 3 |
| Claim | Whether AlphaZero-style self-play discovers the constructions is open |
| Coverage category | Open or deferred item |
| Statement source | paper_jmlr.tex lines 1261–1261; SHA-256 of the normalised block source 8e6dfe8ac49a39a7… |
| Explanations | explain/c-selfplay-open, formal/c-selfplay-open |
| Build evidence | fresh isolated rebuild, audit report |
4 Lean statement and source
No Lean statement corresponds to this claim. This item is explicitly open or deferred.
Build evidence: fresh isolated rebuild of all 38 bundled modules with Lean (version 4.26.0, started 2026-10-10T15:21:56Z, finished 2026-10-10T15:27:00Z: 38/38 modules compiled, 1588 axiom reports, no errors. See Build for the commands and the independent audit of 2026-10-10 (dated evidence).