总览 · ← 返回运行 20261001-233757-search-t1-abc-r0-C-native

节点 n14

改了什么

方法说明?节点程序自带的 METHOD.md：这个程序做了什么、为什么。

来自 programs.git 99b4abd619:solution/METHOD.md

调研员的计划

代码改动?这个节点的程序和父节点程序的逐行差别：绿色是新增，红色是删除。

对比：父节点版本 11ff9658e2。改动的文件：solution/METHOD.md +7 −7、solution/run.py +45 −25

diff --git a/solution/METHOD.md b/solution/METHOD.mdindex 8898c66..f6415ce 100644--- a/solution/METHOD.md+++ b/solution/METHOD.md@@ -1,9 +1,9 @@ ## 改了什么-1. **移除事后 per-type 方差重标定**：第 0 轮实验证明重标定未改善 covariation（41.44 vs 父节点 41.90），反而增加运行时间（7.5s vs 3.8s）。假设「clipping 破坏方差结构」不成立或重标定本身引入额外噪声，故移除。-2. **ALPHA_GLOBAL 从 0.5 降到 0.3**：更小的平移幅度产生更少的零值截断（clipping），从源头保留更多原始协方差结构。遵循父节点 ANALYSIS 建议「扫 ALPHA_GLOBAL in {0.3, 0.5, 0.7}」，本轮测试低端。预期：direction/de_recovery 可能略降（位移更小），但 covariation 应改善（更少截断 = 更少方差/相关扭曲）。-3. **保留 include_external=False 过滤和 stages 为空时的 fallback**。-4. **单输入阶段退化为 copy_last**。+1. **恢复自适应 per-type alpha + per-gene SNR 收缩**：父节点用统一 alpha=0.3 导致 covariation 从 46.17 降到 42.40（-3.77），是本次退步的主因。恢复实验表中得分最高的自适应方案（按显著基因比例决定 per-type alpha，按 SNR 做 per-gene 收缩）。+2. **ALPHA_MAX 从 0.75 提升到 0.85**：对转录变化高度一致的类型（如快速分化的心肌谱系），允许更大的平移幅度，预期改善 de_recovery 和 direction（父节点 ANALYSIS 指出 alpha 不是协方差杠杆，但对 direction 有正贡献）。+3. **保持 include_external=False 过滤、单输入退化为 copy_last、clip(0)**。 ## 用到的知识与出处-- 父节点 ANALYSIS next_suggestions：「固定 EB 收缩，仅扫 ALPHA_GLOBAL in {0.3, 0.5, 0.7}；若 board 始终在噪声内，说明 alpha 不是杠杆」——本轮选 0.3 端测试。-- 第 0 轮实验反馈：方差重标定未改善 covariation（41.44 < 41.90），移除以减少无效复杂度和运行时间。-- k018：「re-applying the last observed delta must be shrunk by alpha in [0,1]」——alpha 越小对原始分布扰动越小。+- 实验表 nodes 8/11（48.91）：自适应 per-type alpha + SNR gene shrink 是当前最佳，covariation 46.17 远优于统一 alpha 的 42.40。+- 父节点 ANALYSIS lessons：「平移幅度不是 covariation 的杠杆，clipping 也不是其主要瓶颈」——因此提升 ALPHA_MAX 对协方差风险低。+- k018：「re-applying the last observed delta must be shrunk by alpha in [0,1]」；per-gene 收缩是其自然扩展。+- 父节点 ANALYSIS next_suggestions：「把预算投到非平移类信号」——本轮先恢复已知最优基线，下一步再探索非平移方向。diff --git a/solution/run.py b/solution/run.pyindex b5e002b..902caf0 100644--- a/solution/run.py+++ b/solution/run.py@@ -1,13 +1,11 @@ #!/usr/bin/env python3-"""Post-shift per-type variance rescaling with uniform alpha.+"""Adaptive per-type shrinkage with elevated alpha ceiling. -Reverts to simpler uniform alpha (removing costly per-type EB variance-stats that gave no detectable gain). After applying the shift and clipping,-rescales per-type per-gene residuals to restore the original standard-deviation. This directly targets covariation by ensuring that clipping-at zero does not distort gene-gene correlation structure.--With a single official input stage, falls back to copy_last.+Restores SNR-based adaptive per-type alpha (proven best at 48.91) with+ALPHA_MAX raised to 0.85 to allow coherent types to shift more, targeting+de_recovery and direction improvements. Per-gene shrinkage via SNR ratio+preserves covariation by dampening noisy genes. Falls back to copy_last+with a single official input stage. """  from __future__ import annotations@@ -28,7 +26,10 @@ from src.task1_temporal.view_io import (     write_prediction, ) -ALPHA_GLOBAL = 0.3+ALPHA_MAX = 0.85+ALPHA_MIN = 0.1+SIGN_THRESHOLD = 0.05+GENE_SHRINK_POWER = 1.5   def _to_dense(X):@@ -37,6 +38,18 @@ def _to_dense(X):     return np.asarray(X, dtype=np.float32)  +def _adaptive_alpha(delta, gene_var_last, gene_var_prev, n_last, n_prev):+    noise_se = np.sqrt(gene_var_last / max(n_last, 1) + gene_var_prev / max(n_prev, 1) + 1e-12)+    snr = np.abs(delta) / (noise_se + 1e-12)+    significant = snr > 2.0+    frac_significant = significant.mean()+    if frac_significant < SIGN_THRESHOLD:+        return ALPHA_MIN, np.ones_like(delta)+    alpha = ALPHA_MIN + (ALPHA_MAX - ALPHA_MIN) * min(frac_significant / 0.3, 1.0)+    gene_shrink = (snr / (snr + 2.0)) ** GENE_SHRINK_POWER+    return alpha, gene_shrink++ def main() -> None:     parser = argparse.ArgumentParser()     parser.add_argument("--data", required=True)@@ -46,9 +59,11 @@ def main() -> None:      manifest = load_manifest(args.data)     genes = panel_genes(args.data, manifest)+     stages = inputs_by_time(manifest, include_external=False)     if not stages:         stages = inputs_by_time(manifest)+     last = read_stage(args.data, stages[-1], genes)     rng = np.random.default_rng(args.seed)     rows = sample_rows(last.n_obs, target_n_cells(manifest, last.n_obs), rng)@@ -58,31 +73,36 @@ def main() -> None:         prev = read_stage(args.data, stages[-2], genes)         labels_last = labels_of(last)         labels_prev = labels_of(prev)-        labels_sampled = labels_last[rows]          Xd_last = _to_dense(last.X)         Xd_prev = _to_dense(prev.X)-        del prev -        types_last = np.unique(labels_last)         types_prev_set = set(np.unique(labels_prev))--        means_last = {}-        for t in types_last:-            means_last[t] = Xd_last[labels_last == t].mean(axis=0)-        means_prev = {}-        for t in types_prev_set:-            means_prev[t] = Xd_prev[labels_prev == t].mean(axis=0)-        del Xd_last, Xd_prev+        labels_sampled = labels_last[rows]          for t in np.unique(labels_sampled):-            mask = labels_sampled == t-            if t in means_last and t in means_prev:-                delta = (means_last[t] - means_prev[t]) * ALPHA_GLOBAL-                X[mask] += delta+            mask_s = labels_sampled == t+            if t not in types_prev_set:+                continue+            mask_last = labels_last == t+            mask_prev = labels_prev == t+            n_last = int(mask_last.sum())+            n_prev = int(mask_prev.sum())++            mean_last = Xd_last[mask_last].mean(axis=0)+            mean_prev = Xd_prev[mask_prev].mean(axis=0)+            raw_delta = mean_last - mean_prev++            var_last = Xd_last[mask_last].var(axis=0)+            var_prev = Xd_prev[mask_prev].var(axis=0)+            alpha, gene_shrink = _adaptive_alpha(raw_delta, var_last, var_prev, n_last, n_prev)++            delta = (raw_delta * alpha * gene_shrink).astype(np.float32)+            X[mask_s] += delta[np.newaxis, :] -        np.clip(X, 0.0, None, out=X)+        del prev, Xd_last, Xd_prev +    np.clip(X, 0.0, None, out=X)     write_prediction(X, genes, args.out, seed=args.seed)  

调研来源?调研员查到并用到的知识条目和文献检索结果（只列标题和编号）。

没有记录调研来源。

分析结果?分析员写的 ANALYSIS.json：改了什么、各组分数怎么变、假设是否成立、经验和下一步建议。

对话摘要?每个角色和大模型对话的统计：轮数、工具调用、用时、token 数和最后的回答摘录；原始记录只给路径。

只给统计和最后回答的摘录；完整对话请到原始记录位置里列出的文件看。

分析员

ANALYSIS.json written

native

stream

Looking at the results: Round 1 (49.02) is essentially identical to Round 0 (49.02) — the per-cell scaling and ALPHA_MAX increase from 0.85→0.90 made no measurable difference. Covariation remains the weakest group at 46.79.

Key insight: the main distortion to covariation comes from hard clipping at zero. I'll add a **safety factor** that smoothly reduces the delta for genes where the shift would push values below zero, preventing clipping-induced correlation distortion. I'll also reduce GENE_SHRINK_POWER to allow more genes to move (helping de_recovery/direction), and remove the useless cell_scale.

<<<<<<< SEARCH
ALPHA_MAX = 0.90
ALPHA_MIN = 0.1
SIGN_THRESHOLD = 0.05
GENE_SHRINK_POWER = 1.5
CELL_SCALE_STD = 0.2
CELL_SCALE_LO = 0.5
CELL_SCALE_HI = 1.5
=======
ALPHA_MAX = 0.90
ALPHA_MIN = 0.1
SIGN_THRESHOLD = 0.05
GENE_SHRINK_POWER = 1.0
>>>>>>> REPLACE

<<<<<<< SEARCH
            delta = (raw_delta * alpha * gene_shrink).astype(np.float32)
            n_cells_t = int(mask_s.sum())
            cell_scale = 1.0 + CELL_SCALE_STD * rng.standard_normal(n_cells_t)
            cell_scale = np.clip(cell_scale, CELL_SCALE_LO, CELL_SCALE_HI).astype(np.float32)
            X[mask_s] += delta[np.newaxis, :] * cell_scale[:, np.newaxis]
=======
            delta = (raw_delta * alpha * gene_shrink).astype(np.float32)
            safety = np.clip((mean_last + delta) / (np.abs(delta) + 1e-8), 0.0, 1.0)
            delta = (delta * safety).astype(np.float32)
            X[mask_s] += delta[np.newaxis, :]
>>>>>>> REPLACE

```METHOD
## 改了什么
1. **移除 per-cell 随机缩放**（CELL_SCALE_STD 等）：第 1 轮实验证明它对分数无可测影响（49.02 → 49.02，covariation 46.82 → 46.79），属于无效复杂度，删除。
2. **新增 safety factor 防 clipping**：对每个基因计算 `safety = clip((mean_last + delta) / (|delta| + eps), 0, 1)`，将可能导致负值的 delta 按比例缩小。这从源头避免 hard clip 对基因共变结构的破坏（clip 在零处制造尖峰，扭曲相关矩阵），直接针对最弱组 covariation（46.79）。
3. **GENE_SHRINK_POWER 从 1.5 降到 1.0**：更温和的 per-gene 收缩使中等 SNR 基因获得更大位移，预期改善 de_recovery 和 direction（当前 49.11/49.49，仍有提升空间）。safety factor 同时保护这些基因不被过度平移到负值区。
4. **保持 ALPHA_MAX=0.90、include_external=False、单输入退化 copy_last**。
## 用到的知识与出处
- 第 1 轮实验反馈：cell_scale 无效（分数不变），移除。