File size: 6,947 Bytes
99aac1c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
from __future__ import annotations

import math
import statistics
from typing import Any


SMALL_BETA = 0.25
LARGE_BETA = 5.0
T_CRITICAL_95 = {
    2: 4.303,
    8: 2.306,
}


def _rmse(rows: list[dict[str, Any]], field: str) -> float:
    errors = [
        float(row[field]) - float(row["fresh_reward"])
        for row in rows
    ]
    return math.sqrt(statistics.fmean(error * error for error in errors))


def evaluate(rows: list[dict[str, Any]]) -> dict[str, Any]:
    small = [row for row in rows if float(row["beta_kl"]) == SMALL_BETA]
    large = [row for row in rows if float(row["beta_kl"]) == LARGE_BETA]
    datasets = sorted({str(row["dataset"]) for row in small})
    seeds = sorted({int(row["seed"]) for row in small})

    best_by_trial: dict[tuple[str, int, str], float] = {}
    for row in small:
        key = (str(row["dataset"]), int(row["seed"]), str(row["method"]))
        best_by_trial[key] = max(
            best_by_trial.get(key, -math.inf),
            float(row["objective"]),
        )

    paired_differences = [
        best_by_trial[(dataset, seed, "SOSMC-ULA")]
        - best_by_trial[(dataset, seed, "ImpDiff")]
        for dataset in datasets
        for seed in seeds
    ]
    paired_mean = statistics.fmean(paired_differences)
    if len(paired_differences) > 1:
        paired_sd = statistics.stdev(paired_differences)
        paired_se = paired_sd / math.sqrt(len(paired_differences))
        t_critical = T_CRITICAL_95[len(paired_differences) - 1]
        paired_ci = [
            paired_mean - t_critical * paired_se,
            paired_mean + t_critical * paired_se,
        ]
    else:
        # A dataset shard reports its observed direction without pretending
        # that one paired comparison supports an uncertainty interval. The
        # aggregate three-dataset checker supplies the inferential result.
        paired_sd = None
        paired_se = None
        t_critical = None
        paired_ci = [paired_mean, paired_mean]

    per_dataset_direction = {
        dataset: statistics.fmean(
            best_by_trial[(dataset, seed, "SOSMC-ULA")]
            - best_by_trial[(dataset, seed, "ImpDiff")]
            for seed in seeds
        )
        > 0.0
        for dataset in datasets
    }

    tracking: dict[str, Any] = {}
    for method, field in (
        ("ImpDiff", "particle_reward"),
        ("SOSMC-ULA", "weighted_particle_reward"),
    ):
        method_rows = [
            row
            for row in small
            if row["method"] == method and int(row["step"]) > 0
        ]
        tracking[method] = {
            "n_evaluation_points": len(method_rows),
            "rmse_particle_vs_fresh": _rmse(method_rows, field),
        }

    tracking_better = (
        tracking["SOSMC-ULA"]["rmse_particle_vs_fresh"]
        < tracking["ImpDiff"]["rmse_particle_vs_fresh"]
    )

    variant_names = sorted(
        set.intersection(
            *[
                set(row.get("truth_grid_sensitivity", {}))
                for row in small
            ]
        )
    )
    variant_directions: dict[str, bool] = {}
    variant_gaps: dict[str, float] = {}
    for variant in variant_names:
        variant_best: dict[tuple[str, int, str], float] = {}
        for row in small:
            key = (
                str(row["dataset"]),
                int(row["seed"]),
                str(row["method"]),
            )
            objective = float(
                row["truth_grid_sensitivity"][variant]["objective"]
            )
            variant_best[key] = max(
                variant_best.get(key, -math.inf), objective
            )
        gaps = [
            variant_best[(dataset, seed, "SOSMC-ULA")]
            - variant_best[(dataset, seed, "ImpDiff")]
            for dataset in datasets
            for seed in seeds
        ]
        variant_gaps[variant] = statistics.fmean(gaps)
        variant_directions[variant] = all(gap > 0.0 for gap in gaps)
    grid_sensitivity_passed = bool(variant_names) and all(
        variant_directions.values()
    )

    large_best: dict[str, float] = {}
    if large:
        for method in ("ImpDiff", "SOSMC-ULA"):
            values = [
                float(row["objective"])
                for row in large
                if row["method"] == method
            ]
            large_best[method] = max(values)
        large_gap: float | None = (
            large_best["SOSMC-ULA"] - large_best["ImpDiff"]
        )
        large_comparable: bool | None = abs(large_gap) <= 0.05
    else:
        large_gap = None
        large_comparable = None

    # Negative control: swapping the algorithm labels must break the
    # pre-registered small-beta directional result.
    reversed_direction = paired_mean < 0.0
    negative_control_failed_as_intended = not reversed_direction

    checks = {
        "small_beta_positive_observed_advantage": paired_mean > 0.0,
        "small_beta_positive_advantage_each_dataset": per_dataset_direction,
        "sosmc_weighted_tracking_rmse_below_impdiff_unweighted": tracking_better,
        "small_beta_positive_across_grid_variants": grid_sensitivity_passed,
        "large_beta_best_objectives_within_0p05": large_comparable,
    }
    small_beta_and_tracking_passed = (
        checks["small_beta_positive_observed_advantage"]
        and all(per_dataset_direction.values())
        and tracking_better
        and grid_sensitivity_passed
        and negative_control_failed_as_intended
    )
    passed = small_beta_and_tracking_passed and large_comparable is True
    return {
        "verdict": "VERIFIED" if passed else "BLOCKED",
        "scope": {
            "datasets": datasets,
            "small_beta": SMALL_BETA,
            "small_beta_seeds": seeds,
            "large_beta": LARGE_BETA,
            "large_beta_dataset": "circles",
        },
        "small_beta_best_objective_paired_difference": {
            "n": len(paired_differences),
            "values": paired_differences,
            "mean": paired_mean,
            "sd": paired_sd,
            "ci95": paired_ci,
            "t_critical": t_critical,
        },
        "per_dataset_positive_mean_advantage": per_dataset_direction,
        "tracking": tracking,
        "grid_sensitivity": {
            "variants": variant_names,
            "sosmc_minus_impdiff_best_objective": variant_gaps,
            "positive_direction": variant_directions,
        },
        "large_beta_control": {
            "run": bool(large),
            "best_objectives": large_best,
            "sosmc_minus_impdiff": large_gap,
            "absolute_comparability_threshold": 0.05,
        },
        "checks": checks,
        "small_beta_and_tracking_passed": small_beta_and_tracking_passed,
        "negative_control": {
            "reversed_label_direction_passed": reversed_direction,
            "failed_as_intended": negative_control_failed_as_intended,
        },
        "passed": passed,
    }