BurnyCoder commited on
Commit
6e557a2
·
verified ·
1 Parent(s): c669ac9

Run pqn/seed1 trained

Browse files
reports/experiments/pqn.md CHANGED
@@ -14,13 +14,15 @@ Metric: success-weighted speed (SWS) and finish rate; intervals are 95 % two-lev
14
 
15
  | Run | Status | Phase | Agent steps | Wall hours | Physics | Commit | Evaluated on |
16
  |---|---|---|---|---|---|---|---|
17
- | `pqn/seed1` | planned | 2-single-seed | — | — | v1 | — | — |
18
  | `pqn/seed2` | not started | 6-seed-sensitivity | — | — | v1 | — | — |
19
  | `pqn/seed3` | not started | 6-seed-sensitivity | — | — | v1 | — | — |
20
 
21
  ## Configuration (logged in each run.json)
22
 
23
- _No run of this group has logged its configuration yet (the experiment files are in `configs/experiments/`)._
 
 
24
 
25
  ## Results
26
 
@@ -32,6 +34,12 @@ _No run of this group has logged its configuration yet (the experiment files are
32
 
33
  Rows from other groups are shown for comparison. Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Training sample: 1,000 tracks; Own training tracks: 1,000 tracks; Out-of-range test: 187 tracks; Campaign (hand-made): 5 tracks; Unfiltered test: 200 tracks.
34
 
 
 
 
 
 
 
35
  ## Learnings
36
 
37
  _No hand-written notes yet (add `reports/notes/pqn.md`)._
 
14
 
15
  | Run | Status | Phase | Agent steps | Wall hours | Physics | Commit | Evaluated on |
16
  |---|---|---|---|---|---|---|---|
17
+ | `pqn/seed1` | trained | 2-single-seed | 1,000,341,504 | 1.10 | v1 | 70a35809b4 | — |
18
  | `pqn/seed2` | not started | 6-seed-sensitivity | — | — | v1 | — | — |
19
  | `pqn/seed3` | not started | 6-seed-sensitivity | — | — | v1 | — | — |
20
 
21
  ## Configuration (logged in each run.json)
22
 
23
+ | Run | total_steps | num_envs | num_steps | lr | gamma | gae_lambda | ent_coef | clip_coef | hidden | action_mode | weight_decay | loaded_tracks | window_k | blind | time_cost | fall_penalty | random_start_prob | max_steps | hint_weight | scripted_starts | training tracks |
24
+ |---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
25
+ | `pqn/seed1` | 1000000000 | 16384 | 32 | 0.00025 | 0.99 | 0.95 | 0.01 | 0.2 | [256, 256] | factored | 0.0 | 4096 | 20 | False | 0.0 | 0.0 | 0.5 | 1500 | 0.0 | False | 48949 |
26
 
27
  ## Results
28
 
 
34
 
35
  Rows from other groups are shown for comparison. Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Training sample: 1,000 tracks; Own training tracks: 1,000 tracks; Out-of-range test: 187 tracks; Campaign (hand-made): 5 tracks; Unfiltered test: 200 tracks.
36
 
37
+ ### Training curves
38
+
39
+ ![Training curves of PQN; one line per seed (seeds 1). Return and finish rate are over training episodes (standing and random starts); validation SWS is the evaluator's score on the validation tracks.](../figures/curves_pqn.png)
40
+
41
+ *Training curves of PQN; one line per seed (seeds 1). Return and finish rate are over training episodes (standing and random starts); validation SWS is the evaluator's score on the validation tracks.*
42
+
43
  ## Learnings
44
 
45
  _No hand-written notes yet (add `reports/notes/pqn.md`)._
reports/figures/curves_pqn.png ADDED

Git LFS Details

  • SHA256: 07fb557b262bdea2ab274adc459d4a503998614903e497c9daf740bc01ddb0ec
  • Pointer size: 130 Bytes
  • Size of remote file: 89.4 kB
reports/figures/curves_pqn.svg ADDED
reports/summary.md CHANGED
@@ -10,11 +10,11 @@ _Generated by `uv run python scripts/07_make_report.py` from the files in `resul
10
  |---|---|---|---|---|---|---|
11
  | baselines | 0 | 0 | 0 | 0 | 15 | 15 |
12
  | 1-first-models | 0 | 0 | 0 | 0 | 2 | 2 |
13
- | 2-single-seed | 22 | 1 | 0 | 0 | 2 | 25 |
14
  | 4-imitation-hybrids | 5 | 0 | 0 | 0 | 0 | 5 |
15
  | 5-hyperparameters | 4 | 0 | 0 | 0 | 0 | 4 |
16
  | 6-seed-sensitivity | 65 | 0 | 0 | 0 | 0 | 65 |
17
- | **all** | 96 | 1 | 0 | 0 | 19 | 116 |
18
 
19
  Statuses: **not started** = in the run matrix, not yet registered; **planned** = registered; code and config pushed before training; **training** = training in progress (or interrupted and resumable); **trained** = trained to its step budget, not yet evaluated; **evaluated** = evaluated; scores available.
20
 
@@ -65,7 +65,7 @@ Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Train
65
  | Blind control | Control | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + nearest centreline point only (blind) | 48,949 generated tracks | standing + random centreline (p=0.5) | 1 | 1 evaluated, 4 not started | Test 1.010 / 73 %; Val 1.003 / 73 %; Train 1.008 / 73 %; Own train 1.008 / 73 %; OOR 0.605 / 44 %; Campaign 1.443 / 100 %; Unfiltered 1.022 / 74 % |
66
  | Generalist (PPO) [pilot, seed 0] | Main method | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + 20 look-ahead points | 48,949 generated tracks | standing + random centreline (p=0.5) | 0 | 1 evaluated | Test 1.199 / 80 %; Val 1.242 / 83 %; Train 1.234 / 83 %; OOR 0.788 / 52 %; Campaign 1.596 / 100 %; Unfiltered 1.227 / 82 % |
67
  | Generalist (PPO) | Main method | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + 20 look-ahead points | 48,949 generated tracks | standing + random centreline (p=0.5) | 1 | 1 evaluated, 9 not started | Test 1.214 / 81 %; Val 1.236 / 83 %; Train 1.250 / 84 %; Own train 1.250 / 84 %; OOR 0.875 / 58 %; Campaign 1.590 / 100 %; Unfiltered 1.231 / 82 % |
68
- | PQN | Baseline algorithm | — | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | generated training pool | standing + random centreline | 1 | 1 planned, 2 not started | — |
69
  | Trained on 1 track | Ablation | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | 1 generated tracks | standing + random centreline | — | 5 not started | — |
70
  | Trained on 10 tracks | Ablation | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | 10 generated tracks | standing + random centreline | — | 3 not started | — |
71
  | Trained on 100 tracks | Ablation | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | 100 generated tracks | standing + random centreline | — | 3 not started | — |
@@ -109,7 +109,7 @@ Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Train
109
  | [Blind control](experiments/blind-control.md) | Does the agent use the look-ahead (H3)? | 1 evaluated, 4 not started |
110
  | [Training-set size](experiments/training-set-size.md) | Is track diversity what generalises (H2)? | 17 not started |
111
  | [Specialists](experiments/specialists.md) | How much time does a generalist lose against single-track specialists? | 30 not started |
112
- | [Baseline algorithm: PQN](experiments/pqn.md) | Does the result depend on PPO? | 1 planned, 2 not started |
113
  | [Ablations: inputs, actions and reward](experiments/ablations.md) | Which design choices matter (exploratory)? | 21 not started |
114
  | [Imitation and hybrids (phase 4)](experiments/imitation-hybrids.md) | What does help from the scripted driver change (exploratory)? | 5 not started |
115
  | [Hyperparameter sensitivity (phase 5)](experiments/hyperparameters.md) | How far does the test score move when one setting changes? | 8 not started |
@@ -132,7 +132,8 @@ Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Train
132
  |---|---|---|
133
  | oor | RandomDriver | 3 |
134
  | oor | blind/seed1 | 1 |
135
- | oor | generalist/seed0 | 1 |
 
136
  | oor | generalist/seed1 | 1 |
137
  | oor | scripted | 1 |
138
  | oor | untrained-seed0 | 1 |
@@ -148,10 +149,10 @@ Track counts: Test (unseen generated): 971 tracks; Validation: 197 tracks; Train
148
  | test | RandomDriver | 3 |
149
  | test | blind/seed1 | 1 |
150
  | test | cem | 2 |
151
- | test | export:unseen-track-demo | 3 |
152
- | test | generalist/seed0 | 1 |
153
  | test | generalist/seed1 | 1 |
154
- | test | scripted | 2 |
155
  | test | untrained-seed0 | 1 |
156
  | test | untrained-seed1 | 1 |
157
  | test | untrained-seed2 | 1 |
 
10
  |---|---|---|---|---|---|---|
11
  | baselines | 0 | 0 | 0 | 0 | 15 | 15 |
12
  | 1-first-models | 0 | 0 | 0 | 0 | 2 | 2 |
13
+ | 2-single-seed | 22 | 0 | 0 | 1 | 2 | 25 |
14
  | 4-imitation-hybrids | 5 | 0 | 0 | 0 | 0 | 5 |
15
  | 5-hyperparameters | 4 | 0 | 0 | 0 | 0 | 4 |
16
  | 6-seed-sensitivity | 65 | 0 | 0 | 0 | 0 | 65 |
17
+ | **all** | 96 | 0 | 0 | 1 | 19 | 116 |
18
 
19
  Statuses: **not started** = in the run matrix, not yet registered; **planned** = registered; code and config pushed before training; **training** = training in progress (or interrupted and resumable); **trained** = trained to its step budget, not yet evaluated; **evaluated** = evaluated; scores available.
20
 
 
65
  | Blind control | Control | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + nearest centreline point only (blind) | 48,949 generated tracks | standing + random centreline (p=0.5) | 1 | 1 evaluated, 4 not started | Test 1.010 / 73 %; Val 1.003 / 73 %; Train 1.008 / 73 %; Own train 1.008 / 73 %; OOR 0.605 / 44 %; Campaign 1.443 / 100 %; Unfiltered 1.022 / 74 % |
66
  | Generalist (PPO) [pilot, seed 0] | Main method | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + 20 look-ahead points | 48,949 generated tracks | standing + random centreline (p=0.5) | 0 | 1 evaluated | Test 1.199 / 80 %; Val 1.242 / 83 %; Train 1.234 / 83 %; OOR 0.788 / 52 %; Campaign 1.596 / 100 %; Unfiltered 1.227 / 82 % |
67
  | Generalist (PPO) | Main method | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + 20 look-ahead points | 48,949 generated tracks | standing + random centreline (p=0.5) | 1 | 1 evaluated, 9 not started | Test 1.214 / 81 %; Val 1.236 / 83 %; Train 1.250 / 84 %; Own train 1.250 / 84 %; OOR 0.875 / 58 %; Campaign 1.590 / 100 %; Unfiltered 1.231 / 82 % |
68
+ | PQN | Baseline algorithm | PQN | Q-network 2x256, LayerNorm, one 9-way head | centreline progress | car state + 20 look-ahead points | 48,949 generated tracks | standing + random centreline (p=0.5) | 1 | 1 trained, 2 not started | — |
69
  | Trained on 1 track | Ablation | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | 1 generated tracks | standing + random centreline | — | 5 not started | — |
70
  | Trained on 10 tracks | Ablation | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | 10 generated tracks | standing + random centreline | — | 3 not started | — |
71
  | Trained on 100 tracks | Ablation | PPO | MLP 2x256 tanh (actor and critic), two 3-way heads | centreline progress | car state + look-ahead points (default) | 100 generated tracks | standing + random centreline | — | 3 not started | — |
 
109
  | [Blind control](experiments/blind-control.md) | Does the agent use the look-ahead (H3)? | 1 evaluated, 4 not started |
110
  | [Training-set size](experiments/training-set-size.md) | Is track diversity what generalises (H2)? | 17 not started |
111
  | [Specialists](experiments/specialists.md) | How much time does a generalist lose against single-track specialists? | 30 not started |
112
+ | [Baseline algorithm: PQN](experiments/pqn.md) | Does the result depend on PPO? | 1 trained, 2 not started |
113
  | [Ablations: inputs, actions and reward](experiments/ablations.md) | Which design choices matter (exploratory)? | 21 not started |
114
  | [Imitation and hybrids (phase 4)](experiments/imitation-hybrids.md) | What does help from the scripted driver change (exploratory)? | 5 not started |
115
  | [Hyperparameter sensitivity (phase 5)](experiments/hyperparameters.md) | How far does the test score move when one setting changes? | 8 not started |
 
132
  |---|---|---|
133
  | oor | RandomDriver | 3 |
134
  | oor | blind/seed1 | 1 |
135
+ | oor | export:unseen-track-demo | 1 |
136
+ | oor | generalist/seed0 | 2 |
137
  | oor | generalist/seed1 | 1 |
138
  | oor | scripted | 1 |
139
  | oor | untrained-seed0 | 1 |
 
149
  | test | RandomDriver | 3 |
150
  | test | blind/seed1 | 1 |
151
  | test | cem | 2 |
152
+ | test | export:unseen-track-demo | 4 |
153
+ | test | generalist/seed0 | 2 |
154
  | test | generalist/seed1 | 1 |
155
+ | test | scripted | 3 |
156
  | test | untrained-seed0 | 1 |
157
  | test | untrained-seed1 | 1 |
158
  | test | untrained-seed2 | 1 |