-
Notifications
You must be signed in to change notification settings - Fork 3
/
params.py
39 lines (28 loc) · 1.13 KB
/
params.py
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
############################# Parameters ###############
#Define input params
batch_size = 25 # batch_size i.e. #(experiences) sampled from exp dataset
im_height = im_width = 80
HISTORY = 3
input_shape = (HISTORY,im_height,im_width)
#replay memory in #(experiences)
EXPERIENCE_SIZE = 50001 #total replay memory dataset
POPULATE = 10000
EPSILON_CHANGE = 1.0005
TARGET_UPDATE_FREQ = 3000 ## When to update target weights i.e. "fixed targets"
SAVE_NETWORK_FREQ = 20000 ## When to save network weights
LEARNING_RATE = 1e-6
#Greedy Approach (Deterministic Network))
INITIAL_EPSILON = .1 #initial value
FINAL_EPSILON = 0.90 #final value
EPSILON_ANNEAL = 1000000
EPSILON_CHANGE = (FINAL_EPSILON - INITIAL_EPSILON) / EPSILON_ANNEAL
#Rewards defined in wrapped_flappy_bird.py
DISCOUNT = 0.95
## TO LOAD QUEUE FROM DATABASE?
LOAD_POPULATED_QUEUE = True
PRETRAINED = False
PRETRAINED_PATH = 'saved nets/weights_iter_000000'
TRAIN_PRETRAINED = True
TRAIN_PRETRAINED_PATH = 'saved nets/weights_iter_420000'
### Epochs when to check average scoreDigitsepochs = [10000,30000,60000,1e5]
SAVE_QUEUE_FREQ = 30000