"""Reporting bias: some sources report a rare sighting more readily than a miss. The bot logs every run, and bulk-pasted logs record every run of the players who kept them ("complete" groups). Form submissions and reports are chosen by the rider, and a run with a shadow or a meadow Blue is more likely to be written up than one without. For each rare roll and each selective group, the bias is the factor by which the group's odds of the rare outcome exceed the complete groups' odds: rho = odds(rare | group) / odds(rare | complete groups) Each rare outcome a group reports then counts 1/rho in the route model's tables. That keeps the group's pattern across routes and moves its overall level to what complete logging shows. The assumption: the bias is the same on every route (planner/policy.yaml says which groups, rolls and settings are used). """ import random from . import model def group_of(r, groups, rare_goals): """The first group in planner/policy.yaml whose condition the run meets.""" for g in groups: w = g.get("when", {}) if "source" in w and r.get("source") != w["source"]: continue if "entry_not" in w and r.get("entry") in (None, w["entry_not"]): continue if w.get("goal") == "rare" and r.get("goal") not in rare_goals: continue return g["name"] return None def _odds(y, n): p = (y + 0.5) / (n + 1) # Jeffreys return p / (1 - p) def estimate(runs, prepared, cfg, rare_goals, seed=20261109, sims=4000): """{group: {roll: {rare, n, rho, lo, hi, used}}} and the group of every run.""" groups = cfg["groups"] names = [group_of(r, groups, rare_goals) for r in runs] complete = {g["name"] for g in groups if g.get("complete")} rolls = cfg["rolls"] tally = {} for name, v in zip(names, prepared): for node, rare in rolls.items(): y = model.OBSERVE[node](v) if y is None: continue t = tally.setdefault(name, {}).setdefault(node, [0, 0]) t[0] += y == rare t[1] += 1 ref = {node: [sum(tally.get(g, {}).get(node, [0, 0])[i] for g in complete) for i in (0, 1)] for node in rolls} rnd = random.Random(seed) out = {} for g in groups: name = g["name"] out[name] = {} for node in rolls: y, n = tally.get(name, {}).get(node, [0, 0]) ry, rn = ref[node] e = {"rare": y, "n": n, "completeRare": ry, "completeN": rn} if g.get("complete"): e.update(rho=1.0, lo=1.0, hi=1.0, used="complete: counted as reported") elif ry < cfg.get("min_complete_rare", 5): # complete logging has (almost) never seen it: the bias cannot be estimated, so this roll is learned # from complete logging only (its range then shows how little is known) e.update(rho=None, lo=None, hi=None, exclude=True, used=f"left out: complete logs have only {ry} rare outcomes, too few to estimate the bias") elif n < cfg.get("min_runs", 30) or rn < cfg.get("min_runs", 30): e.update(rho=None, used=f"too few runs: uses {g.get('same_as') or 'no correction'}") else: rho = _odds(y, n) / _odds(ry, rn) sim = [] # a 90% range for rho from the two rates' posteriors for _ in range(sims): a = rnd.betavariate(y + 0.5, n - y + 0.5) b = rnd.betavariate(ry + 0.5, rn - ry + 0.5) sim.append((a / (1 - a)) / (b / (1 - b))) sim.sort() e.update(rho=rho, lo=sim[int(0.05 * sims)], hi=sim[int(0.95 * sims) - 1], used="estimated") out[name][node] = e # groups with too little data borrow another group's estimate for g in groups: for node in rolls: e = out[g["name"]][node] if e.get("rho") is None and not e.get("exclude"): src = out.get(g.get("same_as") or "", {}).get(node, {}) e["rho"] = src.get("rho") or 1.0 e["lo"], e["hi"] = src.get("lo", e["rho"]), src.get("hi", e["rho"]) return out, names def weights(names, prepared, est, rolls): """Per run: {roll: weight of the outcome it observed}: 1/rho for a rare outcome from a biased group, 0 for any outcome of a roll its group is left out of. Only where it differs from 1.""" out = [] for name, v in zip(names, prepared): w = {} for node, rare in rolls.items(): e = (est.get(name) or {}).get(node, {}) y = model.OBSERVE[node](v) if y is None: continue if e.get("exclude"): w[node] = 0.0 elif y == rare and (e.get("rho") or 1.0) != 1.0: w[node] = 1.0 / e["rho"] out.append(w) return out