Route model learns from bot runs; Thompson sampling in the advisor (work in progress)

- planner/policy.yaml: which sources the route model learns from (spreadsheet,
  reported, bot) and the low-bias sample
- bot runs: deaths coded by stage, Malek intent for runs with a target
- model: survival of the first dragon is not counted when nothing after it is
  known; the meadow dragon falls back to what the bot saw
- solver.js: advisor(..., {sampling: "thompson", seed, draws}) follows one
  posterior draw per attempt and reports each pick's propensity

Not shipped yet: mixing sources lowers the fitted shrinkage for the shadow roll
(alpha 500 -> 50 in the low-bias sample), see the conversation of 2026-11-09.

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 d37eaa3378
commit c0355616b6
6 changed files with 121 additions and 22 deletions
+9
View File
@@ -198,6 +198,15 @@ def bot(name, spec):
goal=target if target in COLOURS else None,
outcome=outcome, outcome_color=colour, data_flags="; ".join(flags), **v)
row["attempt_group"] = attempt_group(row["attempt"])
if target: # the bot never bonds a dragon it was not sent for: it dies on purpose instead (Malek)
row["malek_intent"] = "Yes"
if outcome == "Died before the hour" and not o.get("intentional"): # where the game killed it
if row["scream"] == "Unknown":
row["scream"] = "DIED at first sighting"
elif row["scream"] == "Investigate & help" and row["resolution"] == "Unknown":
row["resolution"] = "DIED in combat"
else:
row["hour"] = "DIED before the hour"
# what the bot saw: the shadow at the treetop, and which dragons waited in the meadow
shadow = _seen_after(steps, 0, "shadow")
if row["hour"] == "Tree" and shadow:
+7 -2
View File
@@ -69,6 +69,9 @@ def prepare_run(r):
v["diedSight"] = r["scream"] == "DIED at first sighting"
if v["diedSight"]:
return v
if r["scream"] not in ("Investigate & help", "Ignore the scream", "Investigate & leave"):
v["sightUnknown"] = True # no answer after the first dragon: whether the rider survived it is not known
return v
res = r["resolution"]
if r["scream"] == "Investigate & help":
v["help"] = True
@@ -113,7 +116,9 @@ def prepare_run(r):
ended = lambda c: r.get("enc_" + c) and r.get("enc_" + c) != "Changed my mind" # noqa: E731
v["meadow"] = "Blue" if (r.get("enc_Blue") or r.get("rare_reported") == "Blue (valley)") else \
"Orange" if ended("Orange") else "Brown" if ended("Brown") else \
"Orange" if r.get("enc_Orange") else "Brown" if r.get("enc_Brown") else None
"Orange" if r.get("enc_Orange") else "Brown" if r.get("enc_Brown") else \
"Blue" if r.get("saw_blue_meadow") is True else \
r["outcome_color"] if r.get("saw_blue_meadow") is False and r.get("outcome_color") in ("Orange", "Brown") else None
if v["meadow"]:
v["encCtx"] = "meadow-" + v["meadow"]
if v.get("encCtx"):
@@ -127,7 +132,7 @@ def prepare_run(r):
# what each chance event observed in a prepared run
OBSERVE = {
"fs": lambda v: v.get("fs"),
"dieSight": lambda v: ("die" if v.get("diedSight") else "live") if v.get("fs") else None,
"dieSight": lambda v: ("die" if v.get("diedSight") else "live") if v.get("fs") and not v.get("sightUnknown") else None,
"dieCombat": lambda v: ("die" if v.get("diedCombat") else "live") if v.get("help") else None,
"dieHour": lambda v: ("die" if v.get("diedHour") else "live") if v.get("survivedCadets") else None,
"shadow": lambda v: None if "shadow" not in v else ("yes" if v["shadow"] else "no"),
+52 -14
View File
@@ -24,8 +24,33 @@
const encColour = (ctx) => (ctx === "tree-Greens" ? "Green" : ctx.split("-")[1]);
const api = {};
/* ---------- the odds: posterior means with hierarchical back-off (as in core/atlas/model.py) ---------- */
function makeParams(T, { alphas = {}, game }) {
/* ---------- posterior draws, for Thompson sampling ---------- */
function rng(seed) {
let x = (seed >>> 0) || 0x9e3779b9;
return () => { x ^= x << 13; x >>>= 0; x ^= x >>> 17; x ^= x << 5; x >>>= 0; return (x + 0.5) / 4294967296; };
}
function normal(rand) { return Math.sqrt(-2 * Math.log(rand())) * Math.cos(2 * Math.PI * rand()); }
function gamma(k, rand) { // Marsaglia and Tsang
if (k < 1) return gamma(k + 1, rand) * Math.pow(rand(), 1 / k);
const d = k - 1 / 3, c = 1 / Math.sqrt(9 * d);
for (;;) {
let x, v;
do { x = normal(rand); v = 1 + c * x; } while (v <= 0);
v = v * v * v;
const u = rand();
if (u < 1 - 0.0331 * x ** 4 || Math.log(u) < 0.5 * x * x + d * (1 - v + Math.log(v))) return d * v;
}
}
function dirichlet(a, rand) {
const g = a.map((x) => Math.max(gamma(x, rand), 1e-300));
let s = 0; for (const x of g) s += x;
return g.map((x) => x / s);
}
/* ---------- the odds: posterior means (or, with rand, one posterior draw) with hierarchical back-off ---------- */
function makeParams(T, { alphas = {}, game, rand = null }) {
const normalise = (a) => { let sum = 0; for (const x of a) sum += x; return a.map((x) => x / sum); };
const estimate = rand ? (a) => dirichlet(a, rand) : normalise;
const memo = new Map();
const ctxKey = (s, names) => names.map((k) => (s[k] == null ? "?" : s[k])).join("|");
function chance(node, state) {
@@ -38,8 +63,7 @@
if (memo.has(mk)) { p = memo.get(mk); continue; }
const c = T[node][key] || {}, al = alphas[node] || ALPHA;
const a = cats.map((k, i) => (c[k] || 0) + (p ? al * p[i] : 0.5));
let sum = 0; for (const x of a) sum += x;
p = a.map((x) => x / sum);
p = estimate(a);
memo.set(mk, p);
}
return Object.fromEntries(cats.map((k, i) => [k, p[i]]));
@@ -49,9 +73,8 @@
if (!memo.has(mk)) {
const e = T.encounter[ctx] || { outcomes: {}, seen: {} };
const cats = game.encounter_outcomes.filter((k) => k === "burn" || e.seen[k] || (e.outcomes[k] || 0) > 0 || k === encColour(ctx));
const a = cats.map((k) => (e.outcomes[k] || 0) + 0.5);
let sum = 0; for (const x of a) sum += x;
memo.set(mk, Object.fromEntries(cats.map((k, i) => [k, a[i] / sum])));
const a = estimate(cats.map((k) => (e.outcomes[k] || 0) + 0.5));
memo.set(mk, Object.fromEntries(cats.map((k, i) => [k, a[i]])));
}
return memo.get(mk);
}
@@ -149,17 +172,31 @@
* the deadly roll just passed on the step after it. When h already has a later step but skips an earlier
* decision (an older form record, say), the recommended option is assumed for the skipped one.
*/
function advisor(sample, X, objective = "eventually", { only = null } = {}) {
function advisor(sample, X, objective = "eventually", { only = null, sampling = "best", seed = 1, draws = 40 } = {}) {
const m = sample.model, game = m.game;
api.game = game; api.ENCOUNTERS = game.by_colour;
const params = makeParams(m.tables, { alphas: m.alphas, game });
const sol = objective === "eventually" ? solveRatio(params, game, m.support, X, only) : solve(params, game, m.support, X, 0, only);
const solveWith = (params) => (objective === "eventually" ? solveRatio(params, game, m.support, X, only) : solve(params, game, m.support, X, 0, only));
// "best": the best route under the posterior means. "thompson": the best route of one posterior draw of the odds
// (one per attempt: make a new advisor with a new seed for each attempt), so every option is tried about as
// often as it is likely to be the best one; as the data firm up, sampling settles on the best route.
const thompson = sampling === "thompson";
const params = makeParams(m.tables, { alphas: m.alphas, game, rand: thompson ? rng(seed) : null });
const sol = solveWith(params);
// the chance Thompson sampling picks each option at a history, estimated from `draws` other draws: the
// propensity to log, so the runs can be re-weighted later
let worlds = null;
const propensityAt = (key, choice) => {
if (!thompson) return 1;
if (!worlds) worlds = Array.from({ length: draws }, (_, i) => solveWith(makeParams(m.tables, { alphas: m.alphas, game, rand: rng(seed * 7919 + i + 1) })).choices);
return worlds.filter((c) => c.get(key) === choice).length / draws;
};
const VARS = new Set([...Object.keys(game.decisions), ...Object.keys(game.chance), "encounter"]);
const fmt = (v) => ({ pX: v.x, pOther: v.o, pReach: v.r });
const label = (k) => (game.decisions[k] || game.chance[k]).label;
const decisionAt = (name, key, extra) => {
const vals = sol.optVals.get(key) || {};
return { kind: "decision", var: name, label: label(name), choice: sol.choices.get(key),
const choice = sol.choices.get(key);
return { kind: "decision", var: name, label: label(name), choice, propensity: propensityAt(key, choice),
options: Object.fromEntries(Object.entries(vals).map(([o, v]) => [o, fmt(v)])), ...extra };
};
@@ -197,13 +234,13 @@
throw new Error("the game has no end");
}
return {
colour: X, objective, only, value: sol.value, step, params, game,
colour: X, objective, only, sampling, seed: thompson ? seed : null, value: sol.value, step, params, game,
/** attempts recorded at an encounter (genuine attempts only), to show how much data a bond chance rests on */
attempts: (ctx) => Object.values(((m.tables.encounter || {})[ctx] || { outcomes: {} }).outcomes).reduce((a, b) => a + b, 0),
/** the next decision only (null when the next step is a roll or the run is over) */
next(h) {
const s = step(h);
return s.kind === "decision" ? { decision: s.var, label: s.label, choice: s.choice, options: s.options } : null;
return s.kind === "decision" ? { decision: s.var, label: s.label, choice: s.choice, propensity: s.propensity, options: s.options } : null;
},
};
}
@@ -222,7 +259,8 @@
const pat = ((adv.game.bot_text || {})[rec.decision] || {})[rec.choice];
const idx = pat ? offered.map(norm).findIndex((t) => new RegExp(pat).test(t)) : -1;
const values = Object.values(rec.options).map((v) => v.pX).sort((a, b) => b - a);
return { index: idx < 0 ? null : idx, decision: rec.decision, choice: rec.choice, margin: values.length > 1 ? values[0] - values[1] : null,
return { index: idx < 0 ? null : idx, decision: rec.decision, choice: rec.choice, propensity: rec.propensity,
margin: values.length > 1 ? values[0] - values[1] : null,
options: rec.options,
reason: idx < 0 && rec.decision !== "encounter" ? "the recommended choice is not on this screen; use the bot's own rules" : `${rec.label}: ${rec.choice}` };
}
+21 -6
View File
@@ -2,7 +2,8 @@
python planner/build_policy.py [--draws 200] [--out data/public/policy.json]
Two samples: every run, and a low-bias sample (bulk logs plus form runs from riders not chasing a rare dragon).
Two samples, from the runs and rules in planner/policy.yaml: every run, and a low-bias sample (bot runs, bulk logs,
and riders not chasing a rare dragon).
The output has a content id, so each bot run can record which policy it followed.
"""
import argparse
@@ -20,9 +21,19 @@ from atlas import runs as runs_io # noqa: E402
import yaml # noqa: E402
# the same runs the site's statistics use (analysis/stats/measures.yaml `sources`)
with open(ROOT / "analysis/stats/measures.yaml", encoding="utf-8") as _f:
SOURCES = yaml.safe_load(_f)["sources"]
# which runs the model learns from, and the low-bias sample (planner/policy.yaml)
with open(ROOT / "planner/policy.yaml", encoding="utf-8") as _f:
CONFIG = yaml.safe_load(_f)
SOURCES = CONFIG["sources"]
def low_bias(r, lb=CONFIG["low_bias"]):
"""Kept in the low-bias sample: every run of an always-kept source, bulk logs, and riders not chasing a rare dragon."""
if r["source"] in lb["always"]:
return True
if r["source"] == "reddit spreadsheet" and "bulk" in lb["entries_kept"] and r["entry"] not in (None, "Form"):
return True
return r["goal"] not in lb["rare_goals"]
def clean(o):
@@ -49,10 +60,14 @@ def main(argv=None):
rows = len(runs)
samples = {
"all": runs,
"lowBias": [r for r in runs if r["entry"] != "Form" or r["goal"] not in model.RARE_GOALS],
"lowBias": [r for r in runs if low_bias(r)],
}
t0 = time.time()
out = {"schema": 1, "source": {"form_rows": rows}, "samples": {}}
by_source = {}
for r in runs:
by_source[r["source"]] = by_source.get(r["source"], 0) + 1
out = {"schema": 1, "source": {"form_rows": by_source.get("reddit spreadsheet", 0), "runs": rows, "by_source": by_source},
"samples": {}}
for name, rs in samples.items():
out["samples"][name] = clean(solver.build_policy(rs, draws=a.draws))
print(f"{name}: {len(rs)} runs, {time.time() - t0:.1f} s")
+17
View File
@@ -0,0 +1,17 @@
# Which runs the route model (data/public/policy.json) learns from. planner/build_policy.py reads this file.
#
# The model estimates the game's rolls (first dragon, deaths, shadow, meadow dragon, bond chances) given the choices
# made before them, so any run that records its choices and what happened can teach it, whoever chose the route.
# Deliberate deaths (Malek, the bot's "no shadow" / "no Blue" burns) are never counted as failed bond attempts.
version: 1
# the `source` values of data/public/runs_clean.csv the model learns from
sources: [reddit spreadsheet, reported, bot]
# Two samples. "all" is every run above. "lowBias" leaves out runs where a rare sighting is more likely to have
# been reported than a miss: form submissions and reports from riders chasing Blue or Black. The bot logs every
# run, and bulk-pasted logs record every run of the players who kept them, so both stay in.
low_bias:
always: [bot] # sources whose every run is kept
entries_kept: [bulk] # spreadsheet rows whose `entry` is not "Form" (bulk-pasted logs)
rare_goals: [Blue, Black, Blue or Black]
+15
View File
@@ -110,3 +110,18 @@ test("solverHistory turns form variables into a history keyed by the game's step
assert.strictEqual(l.leave, "Leave them");
assert.strictEqual(l.red, "Let it pass (two Greens)");
});
test("Thompson sampling: one posterior draw per seed, with the chance of each pick", () => {
const S = require("../../core/solver.js");
const P = require("../../data/public/policy.json");
const best = S.advisor(P.samples.all, "Blue", "eventually").next({});
assert.strictEqual(best.propensity, 1);
const a = S.advisor(P.samples.all, "Blue", "eventually", { sampling: "thompson", seed: 3, draws: 8 });
const b = S.advisor(P.samples.all, "Blue", "eventually", { sampling: "thompson", seed: 3, draws: 8 });
const ra = a.next({}), rb = b.next({});
assert.strictEqual(ra.choice, rb.choice); // same seed, same draw
assert.ok(ra.propensity >= 0 && ra.propensity <= 1);
const picks = new Set();
for (let s = 1; s <= 12; s++) picks.add(S.advisor(P.samples.all, "Blue", "eventually", { sampling: "thompson", seed: s, draws: 1 }).next({}).choice);
assert.ok(picks.size > 1, "different seeds explore different options");
});