Route model: correct for reporting bias

- core/atlas/bias.py: for each rare roll (shadow, meadow Blue) and each source group,
  the factor by which the group over-reports the rare outcome relative to complete
  logging (bot, bulk logs), with a 90% range; rare outcomes from biased groups count
  1/rho; a roll the complete logs have (almost) never shown is learned from them only
- planner/policy.yaml: groups, rolls and thresholds; estimates published in
  policy.json under model.reportingBias
- fit_alpha: a roll seen only one way is pooled fully
- docs: goal accounts follow the best route, research accounts use Thompson sampling
  (decisionMode atlas_thompson)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UCzTd31GphpMXDhQqWrCMQ
This commit is contained in:
ekcSchiano
2026-11-10 02:29:09 +00:00
co-authored by Claude Opus 5.5
parent c0355616b6
commit 2271fe0976
8 changed files with 244 additions and 18 deletions
+13 -4
View File
@@ -311,12 +311,18 @@ def summarise(xs):
# ---------------------------------------------------------------- the published policy
def build_policy(runs, draws=200, seed=20261101, min_support=None, alpha=None, colours=None):
"""The full policy for every colour. runs: coded runs (atlas.coding), as dicts."""
def build_policy(runs, draws=200, seed=20261101, min_support=None, alpha=None, colours=None, bias_config=None):
"""The full policy for every colour. runs: coded runs (data/public/runs_clean.csv), as dicts. bias_config:
planner/policy.yaml `reporting_bias` (None: every run counts as reported)."""
colours = colours or model.COLOURS
min_support = G["min_support"] if min_support is None else min_support
prepared = [model.prepare_run(r) for r in runs]
T = model.count_tables(prepared)
bias_est, w = None, None
if bias_config:
from . import bias
bias_est, names = bias.estimate(runs, prepared, bias_config, set(model.RARE_GOALS))
w = bias.weights(names, prepared, bias_est, bias_config["rolls"])
T = model.count_tables(prepared, w)
support, counts = model.support_from(prepared, min_support)
fitted = {k: model.fit_alpha(T, k) for k in model.CHANCE}
alphas = {k: alpha for k in model.CHANCE} if alpha else {k: f["alpha"] for k, f in fitted.items()}
@@ -325,7 +331,8 @@ def build_policy(runs, draws=200, seed=20261101, min_support=None, alpha=None, c
# Cross-fitting: choose the policy on one random half of the runs and score it on the other half.
r = model.Rng(seed)
fold = [0 if r() < 0.5 else 1 for _ in prepared]
halves = [model.count_tables([v for v, f in zip(prepared, fold) if f == h]) for h in (0, 1)]
halves = [model.count_tables([v for v, f in zip(prepared, fold) if f == h],
[x for x, f in zip(w, fold) if f == h] if w else None) for h in (0, 1)]
half_params = [model.Params(Th, "mean", alphas=alphas) for Th in halves]
targets = {}
@@ -366,6 +373,8 @@ def build_policy(runs, draws=200, seed=20261101, min_support=None, alpha=None, c
"chanceContext": {k: s["context"] for k, s in model.CHANCE.items()},
"decisions": model.DECISIONS, "support": support, "supportCounts": counts,
"tables": T,
# how much each kind of source over-reports each rare roll, and the correction applied (bias.py)
"reportingBias": bias_est,
# the game itself (core/spec/game.yaml), so the advisor in core/solver.js can re-solve from any point
"game": {k: G[k] for k in ("version", "colours", "decisions", "chance", "encounters", "encounter_outcomes",
"by_colour", "start", "steps", "from_form", "bot_text")},