forked from falktan/deepracer
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathnotes_technical.txt
75 lines (59 loc) · 3.91 KB
/
notes_technical.txt
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
ToDo
[x] build solid model -> 2005 / 2006
[x] rebuild solid model with new interface
- try out:
[x] 7 steering angles
[x] cut-corners increase
[x] increase Entropie
[ ] anealing Entropie
[x] longer learning
[ ] reduce discount-rate
Championship Cup @ Las Vegas
Vehicles
----
Accenture_005: (grey) 3 CNN, 1.0 m/s, 7 Angles, 1 Speed (mod from Accenture_003)
Accenture_004: (white) 3 CNN, 1.5 m/s, 7 Angles, 1 Speed (mod from Accenture_101010) -> pretty fast
Accenture_003: (orange) 3 CNN, 2.0 m/s, 5 Angles, 1 Speed (mod from Accenture 101010) -> too fast
Accenture_002: (blue) 3 CNN, 1.0 m/s, 5 Angles, 1 Speed
Accenture_001: (red) 3 CNN, 0.3 m/s, 5 Angles, 1 Speed
Accenture_101010: (purple) 3 CNN, 3 m/s, 5 Angles, 1 Speed -> too fast
Accenture 101010: (purple) 5 CNN, 3 m/s, 5 Angles, 1 Speed -> too fast
Models
----
->> !(8.5)(7 angles;) Model-2113: Accenture_005, discount-rate: 0.5, cut-corners: 0.9, update-each: 5, 20 min.
!(8) (7 angles; good; maybe overfitted a little) Model-2112: Accenture_005, discount-rate: 0.5, cut-corners: 0.8, update-each: 5, 25 min.
(7) (fast but good) Model-2111-001: +10 min.
!(8) (fast but good) Model-2111: Accenture_004, discount-rate: 0.5, cut-corners: 0.8, update-each: 5, 25 min.
(1) (too fast) Model-2110: Accenture_003, discount-rate: 0.5, cut-corners: 0.8, update-each: 5, 25 min.
!(6) (clone does not achieve 100% track completion) Model-2109: Accenture_002, S=0.1, discount-rate: 0.5, cut-corners: 0.9, update-each: 5, 60 min.
(8) (looks good; may be underconverged a little) Model-2108: Accenture_002, discount-rate: 0.5, cut-corners: 0.8, update-each: 5, 15 min.
->> !(9) (looks good; might cut corners too much?) Model-2107: Accenture_002, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 18 min.
!(6) (may be overfitted) Model-2106: Accenture_002, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 30 min.
(3) (slow vehicle) Model-2105: Accenture_001, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 30 min.
(0) (not good because of too fast vehicle) Model-2104: Accenture_101010, Speed: 3.0, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 40 min.
(0) (not good because of too fast vehicle) Model-2103: Accenture 101010, Speed: 3.0, discount-rate: 0.999, cut-corners: 1.0, update-each: 7, 40 min.
(2) (good but old UI) Model-2006: Speed: 3.0, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 15 min
(2) (good but old UI) Model-2005: Speed: 3.0, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 20 min
Model-2003: Speed: 3.0, discount-rate: 0.5, cut-corners: 1.0, update-each: 7, 60 min.
Kronberg
---
submissions
1: Model100
2: Model103
3: Model104 <--- the model used for all runs that made it to the leader-board
4: Model107
5: Model108
6: ? didnt work when trying Model104-002
7: ? didnt work when trying again Model104-002
Model104: Speed: 1.5, discount-rate: 0.5, cut-corners: 1.5, update-each:10, 20 min.
Modell05: Speed: 6.0, discount-rate: 0.5, cut-corners: 1.3, update-each:10, 15 min.
Modell06: Speed: 1.5, discount-rate: 0.5, cut-corners: 1.3, max-steering: 20, update-each:10, duration: 15 min.
Model107: Speed: 1.5, discount-rate: 0.5, cut-corners: 1.5, max-steering: 20, steering-df: 3, update-each: 5, duration: 18 min.
Model108: Speed: 1.5, discount-rate: 0.5, cut-corners: 1.5, max-steering: 15, steering-df: 3, update-each: 5, duration: 18 min.
Model121: Speed: 1.5, discount-rate: 0.5, cut-corners: 1.3, update-each:10, 20 min. on London Loop
model104-002: Speed: 1.5, discount-rate: 0.5, cut-corners: 1.5, update-each:10, 20 min. Re-Invent + 12 min. London
Meaning of the parameters
---
- parameters not mentioned are set to the default values
- cut-corners: The factor used to determine the radius of the circle used to find the point the car should aim for. The Radius is the product of this and the width of the track
- update-each: Number of epochs per learning cycle (the last in the list of hyper parameters)