Feasibility-Ranked Group Policy Gradient / results.json
Unverified
1{
2 "toy_verification": {
3 "normalization_sweep": [
4 {
5 "scale": 0.05,
6 "observed_variance": 0.9999996021912955,
7 "predicted": 0.9999996021912961
8 },
9 {
10 "scale": 0.2,
11 "observed_variance": 0.9999999005478004,
12 "predicted": 0.9999999005478019
13 },
14 {
15 "scale": 1.0,
16 "observed_variance": 0.9999999801095585,
17 "predicted": 0.9999999801095593
18 },
19 {
20 "scale": 5.0,
21 "observed_variance": 0.9999999960219117,
22 "predicted": 0.9999999960219117
23 },
24 {
25 "scale": 20.0,
26 "observed_variance": 0.9999999990054788,
27 "predicted": 0.999999999005478
28 }
29 ],
30 "margin_sweep": [
31 {
32 "separation": 0.0,
33 "wF": 1.0,
34 "wV": 1.0,
35 "observed_margin": -0.009074307857396793,
36 "predicted_margin": 0.0
37 },
38 {
39 "separation": 0.0,
40 "wF": 2.0,
41 "wV": 1.0,
42 "observed_margin": -0.013611461786095192,
43 "predicted_margin": 0.0
44 },
45 {
46 "separation": 0.0,
47 "wF": 4.0,
48 "wV": 1.0,
49 "observed_margin": -0.02268576964349199,
50 "predicted_margin": 0.0
51 },
52 {
53 "separation": 0.2,
54 "wF": 1.0,
55 "wV": 1.0,
56 "observed_margin": 0.7356044541460376,
57 "predicted_margin": 0.7440645962550935
58 },
59 {
60 "separation": 0.2,
61 "wF": 2.0,
62 "wV": 1.0,
63 "observed_margin": 1.1034066812190564,
64 "predicted_margin": 1.1160968943826401
65 },
66 {
67 "separation": 0.2,
68 "wF": 4.0,
69 "wV": 1.0,
70 "observed_margin": 1.839011135365094,
71 "predicted_margin": 1.8601614906377335
72 },
73 {
74 "separation": 0.5,
75 "wF": 1.0,
76 "wV": 1.0,
77 "observed_margin": 1.4111123653624242,
78 "predicted_margin": 1.4175707054694646
79 },
80 {
81 "separation": 0.5,
82 "wF": 2.0,
83 "wV": 1.0,
84 "observed_margin": 2.1166685480436365,
85 "predicted_margin": 2.126356058204197
86 },
87 {
88 "separation": 0.5,
89 "wF": 4.0,
90 "wV": 1.0,
91 "observed_margin": 3.5277809134060605,
92 "predicted_margin": 3.5439267636736616
93 },
94 {
95 "separation": 1.0,
96 "wF": 1.0,
97 "wV": 1.0,
98 "observed_margin": 1.788105220470278,
99 "predicted_margin": 1.7921859160787397
100 },
101 {
102 "separation": 1.0,
103 "wF": 2.0,
104 "wV": 1.0,
105 "observed_margin": 2.682157830705417,
106 "predicted_margin": 2.68827887411811
107 },
108 {
109 "separation": 1.0,
110 "wF": 4.0,
111 "wV": 1.0,
112 "observed_margin": 4.470263051175695,
113 "predicted_margin": 4.48046479019685
114 },
115 {
116 "separation": 2.0,
117 "wF": 1.0,
118 "wV": 1.0,
119 "observed_margin": 1.9401772463439695,
120 "predicted_margin": 1.9423874121784488
121 },
122 {
123 "separation": 2.0,
124 "wF": 2.0,
125 "wV": 1.0,
126 "observed_margin": 2.9102658695159542,
127 "predicted_margin": 2.913581118267673
128 },
129 {
130 "separation": 2.0,
131 "wF": 4.0,
132 "wV": 1.0,
133 "observed_margin": 4.850443115859925,
134 "predicted_margin": 4.8559685304461215
135 }
136 ],
137 "margin_observed_slope_wF2_wV1": 1.2815550367972013,
138 "margin_prediction_slope_note": "(wF+wV)/(2*sigma), with sigma measured per sweep"
139 },
140 "mini_experiment": {
141 "device": "cuda",
142 "rows": [
143 {
144 "method": "standard",
145 "final_feasibility_mean": 2.2039625188522364e-09,
146 "final_feasibility_std": 2.464394918026189e-10,
147 "policy_risky_probability_mean": 0.9814547970890999,
148 "expected_reward_mean": 5.9629095941782,
149 "training_rollout_feasibility_last20": 0.0
150 },
151 {
152 "method": "ranked",
153 "final_feasibility_mean": 3.211668065484117e-05,
154 "final_feasibility_std": 6.337542572511882e-06,
155 "policy_risky_probability_mean": 0.8740648701786995,
156 "expected_reward_mean": 5.748129740357399,
157 "training_rollout_feasibility_last20": 9.765625e-05
158 }
159 ]
160 }
161}