Noise-Whitened Trajectory-KL Policy Regularization / bench_report.json
Failed on benchmark
1{
2 "bench_version": 1,
3 "track": "dynamics",
4 "model": "rnn_small",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.01,
10 "reg": 0.03
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "reg": 0.03
17 },
18 "mean": 0.000516663167218212
19 },
20 {
21 "cfg": {
22 "lr": 0.001,
23 "reg": 0.1
24 },
25 "mean": 0.0005222226318437606
26 },
27 {
28 "cfg": {
29 "lr": 0.001,
30 "reg": 0.3
31 },
32 "mean": 0.0005512491334229708
33 },
34 {
35 "cfg": {
36 "lr": 0.003,
37 "reg": 0.03
38 },
39 "mean": 5.068586688139476e-05
40 },
41 {
42 "cfg": {
43 "lr": 0.003,
44 "reg": 0.1
45 },
46 "mean": 5.551056983676972e-05
47 },
48 {
49 "cfg": {
50 "lr": 0.003,
51 "reg": 0.3
52 },
53 "mean": 8.239948328991886e-05
54 },
55 {
56 "cfg": {
57 "lr": 0.01,
58 "reg": 0.03
59 },
60 "mean": 1.4466989341599401e-05
61 },
62 {
63 "cfg": {
64 "lr": 0.01,
65 "reg": 0.1
66 },
67 "mean": 1.7228220713150222e-05
68 },
69 {
70 "cfg": {
71 "lr": 0.01,
72 "reg": 0.3
73 },
74 "mean": 3.9540235775348265e-05
75 }
76 ],
77 "full": {
78 "mean": 1.3864131801710755e-05,
79 "std": 2.917053139718998e-06,
80 "per_seed": [
81 1.2164169675088488e-05,
82 1.9545861505321227e-05,
83 1.0918912266788539e-05,
84 1.5239013919199351e-05,
85 1.629439248063136e-05,
86 1.428560062777251e-05,
87 1.23286408779677e-05,
88 1.0136463060916867e-05
89 ],
90 "n": 8
91 }
92 },
93 "idea": {
94 "mean": 1.4849264971417142e-05,
95 "std": 2.2583458302260994e-06,
96 "per_seed": [
97 1.3598252735391725e-05,
98 1.7523587303003296e-05,
99 1.194078413391253e-05,
100 1.765600791259203e-05,
101 1.7466716599301435e-05,
102 1.4715534234710503e-05,
103 1.3851561561750714e-05,
104 1.2041675290674902e-05
105 ],
106 "n": 8
107 },
108 "comparison": {
109 "delta_mean": 9.85133169706387e-07,
110 "idea_wins": 1,
111 "n_pairs": 8,
112 "per_seed_diffs": [
113 1.4340830603032373e-06,
114 -2.0222742023179308e-06,
115 1.0218718671239913e-06,
116 2.41699399339268e-06,
117 1.1723241186700761e-06,
118 4.299336069379933e-07,
119 1.5229206837830134e-06,
120 1.9052122297580354e-06
121 ],
122 "p_value": 0.0956,
123 "mde": 1.1296557285196537e-06,
124 "mde_rel_pct": 8.148045219681629,
125 "verdict": "no significant win",
126 "system_worked": false
127 },
128 "mechanism_signature": {
129 "selected_cfg": {
130 "lr": 0.01,
131 "reg": 0.03
132 },
133 "idea_grid": [
134 {
135 "cfg": {
136 "lr": 0.001,
137 "reg": 0.03
138 },
139 "result": {
140 "mean": 0.00052486258209683,
141 "std": 5.698033059267576e-05,
142 "per_seed": [
143 0.00045051053166389465,
144 0.0006266050622798502,
145 0.0004562842077575624,
146 0.0005525902379304171,
147 0.0005077053210698068,
148 0.0005103448056615889,
149 0.0005080681294202805,
150 0.0005867923609912395
151 ],
152 "n": 8
153 }
154 },
155 {
156 "cfg": {
157 "lr": 0.001,
158 "reg": 0.1
159 },
160 "result": {
161 "mean": 0.0005531014503503684,
162 "std": 6.433722525463851e-05,
163 "per_seed": [
164 0.0004721346776932478,
165 0.00065583735704422,
166 0.00045898332609795034,
167 0.0005830824957229197,
168 0.0005418244400061667,
169 0.000540129782166332,
170 0.0005432094330899417,
171 0.0006296100909821689
172 ],
173 "n": 8
174 }
175 },
176 {
177 "cfg": {
178 "lr": 0.001,
179 "reg": 0.3
180 },
181 "result": {
182 "mean": 0.0006540729009429924,
183 "std": 8.133966965134085e-05,
184 "per_seed": [
185 0.0005586238112300634,
186 0.0007529730210080743,
187 0.0005105234449729323,
188 0.0006941424799151719,
189 0.0006594074657186866,
190 0.0006338193779811263,
191 0.000662658188957721,
192 0.0007604354177601635
193 ],
194 "n": 8
195 }
196 },
197 {
198 "cfg": {
199 "lr": 0.003,
200 "reg": 0.03
201 },
202 "result": {
203 "mean": 5.004885133530479e-05,
204 "std": 9.772504653528356e-06,
205 "per_seed": [
206 6.087642759666778e-05,
207 6.89755252096802e-05,
208 4.271534999134019e-05,
209 3.5638498957268894e-05,
210 4.9467758799437433e-05,
211 4.593151970766485e-05,
212 4.691871799877845e-05,
213 4.986701242160052e-05
214 ],
215 "n": 8
216 }
217 },
218 {
219 "cfg": {
220 "lr": 0.003,
221 "reg": 0.1
222 },
223 "result": {
224 "mean": 6.530662221848615e-05,
225 "std": 1.2555145162208372e-05,
226 "per_seed": [
227 7.812787953298539e-05,
228 8.879921369953081e-05,
229 4.93601146445144e-05,
230 4.9300262617180124e-05,
231 6.575142469955608e-05,
232 6.285968993324786e-05,
233 6.057017526472919e-05,
234 6.768421735614538e-05
235 ],
236 "n": 8
237 }
238 },
239 {
240 "cfg": {
241 "lr": 0.003,
242 "reg": 0.3
243 },
244 "result": {
245 "mean": 0.0001406980663887225,
246 "std": 2.26067156523767e-05,
247 "per_seed": [
248 0.00016091138240881264,
249 0.0001822679041652009,
250 0.00010793711408041418,
251 0.00011437860666774213,
252 0.00014717398153152317,
253 0.00013510147982742637,
254 0.00013161970127839595,
255 0.00014619436115026474
256 ],
257 "n": 8
258 }
259 },
260 {
261 "cfg": {
262 "lr": 0.01,
263 "reg": 0.03
264 },
265 "result": {
266 "mean": 1.4849264971417142e-05,
267 "std": 2.2583458302260994e-06,
268 "per_seed": [
269 1.3598252735391725e-05,
270 1.7523587303003296e-05,
271 1.194078413391253e-05,
272 1.765600791259203e-05,
273 1.7466716599301435e-05,
274 1.4715534234710503e-05,
275 1.3851561561750714e-05,
276 1.2041675290674902e-05
277 ],
278 "n": 8
279 }
280 },
281 {
282 "cfg": {
283 "lr": 0.01,
284 "reg": 0.1
285 },
286 "result": {
287 "mean": 2.5755787191883428e-05,
288 "std": 4.61634426197786e-06,
289 "per_seed": [
290 2.8114078304497525e-05,
291 1.8590118997963145e-05,
292 2.2178555809659883e-05,
293 3.419041968300007e-05,
294 2.8991913495701738e-05,
295 2.144245991075877e-05,
296 2.614234108477831e-05,
297 2.639641024870798e-05
298 ],
299 "n": 8
300 }
301 },
302 {
303 "cfg": {
304 "lr": 0.01,
305 "reg": 0.3
306 },
307 "result": {
308 "mean": 9.178056825476233e-05,
309 "std": 1.33129719536729e-05,
310 "per_seed": [
311 0.00011019065277650952,
312 7.42549600545317e-05,
313 7.907469989731908e-05,
314 0.00011509206524351612,
315 8.817380148684606e-05,
316 8.460405661026016e-05,
317 9.247064735973254e-05,
318 9.038366260938346e-05
319 ],
320 "n": 8
321 }
322 }
323 ],
324 "signature": {
325 "prediction": "inverse covariance weights low-noise drift more than high-noise drift",
326 "low_high_weighted_cost_ratio_observed": 0.8153622015675527,
327 "models": [
328 {
329 "kind": "baseline",
330 "raw_drift_cost": 0.0008920590626075864,
331 "whitened_drift_cost": 0.002272645477205515
332 },
333 {
334 "kind": "idea",
335 "raw_drift_cost": 0.000797593267634511,
336 "whitened_drift_cost": 0.00203404831700027
337 }
338 ],
339 "confirmed": false
340 },
341 "track_justification": "Dynamics is structurally matched: controlled pendulum rollout forecasting and drift/control regularization."
342 }
343}