Repository navigation
About parameters & bonus #7
Description
we suggest you to use the following parameters to do training
if you think original parameters (described in SPEC) is not good for you
# ============================ Hyper Parameters ============================ #
EPISODE = 500 # training episode
BATCH_SIZE = 128 # batch size
GAMMA = 0.99 # reward discount
TARGET_UPDATE_ITER = 100 # target update frequency
MEMORY_CAPACITY = 2000 # replay buffer size
INIT_REPLAY_SIZE = 1000 # init replay buffer size
RECORD_VIDEO_EVERY = 100
MONITOR_PATH = './vedio' # video path
RECORD_VIDEO_FLAG = False # record video or not
# ============================ Hyper Parameters ============================ #
self.epsilon = 1.0
self.epsilon_min = 0.001
self.epsilon_decay = 0.999
self.learning_rate = 0.001
And We suggest you change the game CartPole-v0 to CartPole-v1 ,
then see the performance between DQN and Double DQN
Additionally, you can modify your network's architecture and describe it in report.
for example:
def __init__(self, ):
super(Net, self).__init__()
self.fc1 = nn.Linear(N_STATES, 32)
self.fc2 = nn.Linear(32, 32)
self.out = nn.Linear(32, N_ACTIONS)
Activity
BIGBALLON commentedon Dec 22, 2017
The results for CartPole-v0 & CartPole-v1
CartPole-v0
CartPole-v1
yyrkoon27 commentedon Dec 22, 2017
Dear TA,
What kind of network architecture and hyper-parameters do you use when generating the v0 and v1 results? They climb up to full reward so fast!
BIGBALLON commentedon Dec 22, 2017
Hi, @yyrkoon27
all of the hyper-parameters and network architecture I used are described in this Issue.
See the details in the first comment.