logaritmisk/trueskill-tt
1
0
Fork 0
Code Issues 1 Pull Requests Packages Projects Releases Activity
140 Commits 1 Branch 13 Tags
bb2a845882326b5565e29c5adf45799077200474
Commit Graph
5 Commits
Author SHA1 Message Date
logaritmiskandClaude Opus 5 bb2a845882 feat!: N-team outcome prediction with draw mass, replacing the 2-team panic 
`predict_outcome` asserted `teams.len() == 2` and returned `[p, 1 - p]`,
allocating no probability to a draw even with `p_draw > 0`. For a
draw-enabled model the numbers were simply wrong, at any team count.

It now returns `Result<Prediction, InferenceError>` and supports N teams.

Two algorithms, both deterministic:

- Who finishes first. Performances are independent Gaussians, so this
  separates into a one-dimensional integral per team rather than a
  multivariate orthant probability. Adaptive Gauss-Kronrod evaluates it
  to ~1e-15, matching the exact two-team closed form.
- A specific finishing order. The factor graph only constrains
  rank-adjacent teams, so a full order is a chain of local constraints,
  not a general orthant integral. That chain collapses into a sequential
  recursion over cumulative integrals: O(teams * grid) per order.

Fixed-node Gauss-Hermite is the obvious tool for the first and is a trap:
when a rival's sigma is small the CDF product becomes a step narrower
than the node spacing, and the nodes step over it. Measured 4.4e-4 off
the closed form on a mildly skewed matchup and 1.7e-2 on a small-sigma
one, while still returning something that looks like a probability.
Adaptive refinement is what makes that case safe, and
`win_probabilities_survive_a_rival_with_a_tiny_sigma` pins it down.

The acceptance test is an identity rather than a golden: the outcome
space is exhaustive and disjoint, so the probabilities sum to one. Any
drift is integration error and nothing else. Gauss-Hermite failed it at
4.4e-4; this holds to ~1e-9.

Also from #21: unknown keys are now reported rather than dropped, so a
team of strangers can no longer produce a confident-looking prediction.
`predict_quality` returns `Result` for the same reason.

BREAKING CHANGE: `predict_outcome` returns `Result<Prediction, _>`
instead of `Vec<f64>`; `predict_quality` returns `Result<f64, _>`.

Refs #21, #39

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011hcFjNDmHXZF8URGLku5zZ
 2026-09-07 14:55:11 +02:00
logaritmiskandClaude Opus 5 1a88678384 refactor!: remove ConvergenceReport::slices_skipped 
Closes #33. The field was public, hardcoded to 0 at both construction sites,
and had no route to ever being non-zero.

It was added in T3 as the reporting surface for dirty-bit slice skipping. That
feature is #4, closed as unworkable — the ceiling measured at ~6% against a
projected 5-50x, on top of three independent soundness blockers. #32, which
reattributed the cost to ingestion, is closed too: the re-convergence is
necessary work rather than waste, because appending one event genuinely moves
the involved competitors ~1.2 sigma across their whole history. Nothing left
would ever populate it.

This is the same defect class as #19, where this arc started: a public surface
that looks implemented, reports a plausible value, and is inert. A caller
reading `slices_skipped: 0` reasonably concludes "no slices were skipped this
run", not "this feature does not exist".

Removed rather than documented as reserved. Its only value was as a hook for a
plan that no longer exists, and keeping it preserves the shape of that plan.
Breaking, but ConvergenceReport is returned rather than constructed by callers,
so the only breakage is code reading a constant zero.

Also added a test asserting every remaining field carries real information —
iterations non-zero, final_step finite, log_evidence a finite negative log
probability, and per_iteration_time holding one duration per iteration.
Mutation-proved: pinning per_iteration_time to an empty SmallVec fails it. The
next always-constant member now has to survive an assertion rather than just a
reviewer's attention, which is the actual lesson of #19 and #33.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T5SYDExxL4vZgvunrcNSMc
 2026-08-28 08:07:41 +02:00
logaritmisk 0705986929 feat(game): plumb ConvergenceOptions through to run_chain 
Game and OwnedGame gain a convergence: ConvergenceOptions field set at
construction. Game::{ranked,scored} forward options.convergence into
OwnedGame::{new,new_scored} (previously dropped on the floor).
{ranked,scored}_with_arena take it as a parameter. run_chain reads
self.convergence.{epsilon, max_iter, alpha} instead of hardcoded
1e-6 / 10 / undamped. DiffFactor::propagate gains an alpha parameter
and dispatches into Trunc/MarginFactor::propagate_with_alpha.

In-tree callsites in src/time_slice.rs and src/history.rs pass
ConvergenceOptions::default(). Pre-existing T2 fallout in tests,
benches, and the atp example (struct literals missing the new alpha
field) is fixed by adding alpha: 1.0 so the workspace builds clean.
Default alpha is 1.0, so all 96 lib + 27 integration test goldens
remain bit-equal.
 2026-05-08 15:10:35 +02:00
logaritmiskandClaude Opus 4.7 8b53cacd64 T4 (MarginFactor): scored outcomes via Gaussian-margin EP evidence 
Adds soft Gaussian-observation evidence on the per-pair diff variable,
enabling continuous score margins as a richer alternative to ranks.

Public API:
- `Outcome::Scored([scores])` (non-breaking enum extension under
  `#[non_exhaustive]`).
- `Game::scored(teams, outcome, options)` constructor parallel to
  `Game::ranked`.
- `EventBuilder::scores([...])` fluent helper.
- `HistoryBuilder::score_sigma(σ)` knob (default 1.0, validated > 0).
- `GameOptions::score_sigma`.
- `EventKind` re-exported from `lib.rs` (annotated `#[non_exhaustive]`).
- New `InferenceError::InvalidParameter { name, value }` variant.

Internals:
- `MarginFactor` (`factor/margin.rs`): Gaussian observation factor that
  closes in one EP step; cavity-cached log-evidence mirrors `TruncFactor`.
- `BuiltinFactor::Margin` dispatch arm.
- `DiffFactor` enum in `game.rs` lets `Game::likelihoods` and the new
  `likelihoods_scored` share the per-pair link abstraction.
- Per-event `EventKind { Ranked, Scored { score_sigma } }` routed through
  `TimeSlice::add_events`, `iteration_direct`, and `log_evidence`.

Tests: 88 lib + 27 integration (4 new in `tests/scored.rs`); existing
goldens byte-identical.  Bench: `benches/scored.rs` baseline ~960µs for
60 events × 20-player pool with default convergence.

Plan: docs/superpowers/plans/2026-04-27-t4-margin-factor.md
Spec item marked Done.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
 2026-04-27 08:47:36 +02:00
logaritmisk d2aab82c1e T0 + T1 + T2: engine redesign through new API surface (#1) 
Implements tiers T0, T1, T2 of `docs/superpowers/specs/2026-04-23-trueskill-engine-redesign-design.md`. All three tiers have landed together on this branch because they build on one another; this PR rolls them up for a single review pass.

Per-tier plans:
- T0: `docs/superpowers/plans/2026-04-23-t0-numerical-parity.md`
- T1: `docs/superpowers/plans/2026-04-24-t1-factor-graph.md`
- T2: `docs/superpowers/plans/2026-04-24-t2-new-api-surface.md`

## Summary

### T0 — Numerical parity (internal)

- `Gaussian` switched to natural-parameter storage `(pi, tau)`; mul/div now ~7× faster (218 ps vs 1.57 ns).
- `HashMap<Index, _>` → dense `Vec<_>` keyed by `Index.0` (via `AgentStore<D>`, `SkillStore`).
- `ScratchArena` eliminates per-event allocations in `Game::likelihoods`.
- `InferenceError` seed type added (1 variant).
- 38 → 53 tests passing through T1.
- Benchmark: `Batch::iteration` 29.84 → 21.25 µs.

### T1 — Factor graph machinery (internal)

- `Factor` trait + `BuiltinFactor` enum (TeamSum / RankDiff / Trunc) driving within-game inference.
- `VarStore` flat storage for variable marginals.
- `Schedule` trait + `EpsilonOrMax` impl replacing the hand-rolled EP loop.
- `Game::likelihoods` rebuilt on the factor-graph machinery; iteration counts and goldens preserved to within 1e-6.
- 53 tests passing.
- Benchmark: `Batch::iteration` 23.01 µs (slight regression absorbed in T2).

### T2 — New API surface (breaking)

**Renames:**
- `IndexMap → KeyTable`, `Player → Rating`, `Agent → Competitor`, `Batch → TimeSlice`

**New types:**
- `Time` trait with `Untimed` ZST and `i64` impls; `Drift<T>`, `Rating<T, D>`, `Competitor<T, D>`, `TimeSlice<T>`, `History<T, D, O, K>` all generic.
- `Event<T, K>`, `Team<K>`, `Member<K>`, `Outcome` (`Ranked` variant; `#[non_exhaustive]`).
- `Observer<T>` trait + `NullObserver`.
- `ConvergenceOptions`, `ConvergenceReport`.
- `GameOptions`, `OwnedGame<T, D>`.

**Three-tier ingestion:**
- `history.record_winner(&K, &K, T)` / `record_draw(&K, &K, T)` — 1v1 convenience.
- `history.add_events(iter)` — typed bulk.
- `history.event(T).team([...]).weights([...]).ranking([...]).commit()` — fluent.

**Query API:** `current_skill`, `learning_curve`, `learning_curves` (keyed on `K`), `log_evidence`, `log_evidence_for`, `predict_quality`, `predict_outcome`.

**Game constructors:** `ranked`, `one_v_one`, `free_for_all`, `custom` — all returning `Result<_, InferenceError>`.

**`factors` module:** `Factor`, `Schedule`, `VarStore`, `VarId`, `BuiltinFactor`, `EpsilonOrMax`, `ScheduleReport`, `TeamSumFactor`, `RankDiffFactor`, `TruncFactor` now public.

**Errors:** `InferenceError` gains `MismatchedShape`, `InvalidProbability`, `ConvergenceFailed`; boundary panics converted to `Result`.

**Removed (breaking):** `History::convergence(iters, eps, verbose)`, `HistoryBuilder::gamma(f64)`, `HistoryBuilder::time(bool)`, `History.time: bool`, `learning_curves_by_index`, nested-Vec public `add_events`.

## Behavior change (documented in CHANGELOG)

`Time = Untimed` has `elapsed_to → 0`, so no drift accumulates between slices. The old `time=false` mode implicitly forced `elapsed=1` on reappearance via an `i64::MAX` sentinel — that quirk is not reproducible under a typed time axis. Tests that depended on it now use `History::<i64, _>` with explicit `1..=n` timestamps. One test (`test_env_ttt`) had 3 Gaussian goldens updated to reflect the corrected semantics; documented in commit `33a7d90`.

## Final numbers

| Metric | Before T0 | After T2 | Delta |
|---|---|---|---|
| `Batch::iteration` | 29.84 µs | 21.36 µs | **-28%** |
| `Gaussian::mul` | 1.57 ns | 219 ps | **-86%** |
| `Gaussian::div` | 1.57 ns | 219 ps | **-86%** |
| Tests passing | 38 | 90 | +52 |

All other Gaussian ops unchanged (~219 ps add/sub, ~264 ps pi/tau reads).

## Test plan

- [x] `cargo test --features approx` — 90/90 pass (68 lib + 10 api_shape + 6 game + 4 record_winner + 2 equivalence)
- [x] `cargo clippy --all-targets --features approx -- -D warnings` — clean
- [x] `cargo +nightly fmt --check` — clean
- [x] `cargo bench --bench batch` — 21.36 µs
- [x] `cargo bench --bench gaussian` — unchanged from T1
- [x] `cargo run --example atp --features approx` — rewritten in new API, runs clean
- [x] Historical Game-level goldens preserved in `tests/equivalence.rs`
- [x] Public API matches spec Section 4 (verified by integration tests in `tests/api_shape.rs`)

## Commit history

~45 commits total across T0 + T1 + T2. Each task is self-contained and individually tested; the branch is bisectable. See `git log main..t2-new-api-surface` for the full list.

## Deferred to later tiers

- `Outcome::Scored` + `MarginFactor` — T4
- `Damped` / `Residual` schedules — T4
- `Send + Sync` bounds + Rayon parallelism — T3
- N-team `predict_outcome` — T4
- `Game::custom` full ergonomics — T4

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Reviewed-on: #1
Co-authored-by: Anders Olsson <anders.e.olsson@gmail.com>
Co-committed-by: Anders Olsson <anders.e.olsson@gmail.com>
 2026-04-24 11:20:04 +00:00