DROO demo_on

完整代码在文章作者的仓库：https://github.com/revenol/DROO。为方便自己看代码所以放上来，看完就删。
这个是论文《Deep Reinforcement Learning for Online Ofﬂoading in Wireless Powered Mobile-Edge Computing Networks》的tf1.x版本代码。

#  #################################################################
#  Deep Reinforcement Learning for Online Ofﬂoading in Wireless Powered Mobile-Edge Computing Networks
#
#  This file contains a demo evaluating the performance of DROO by randomly turning on/off some WDs. It loads the training samples from ./data/data_#.mat, where # denotes the number of active WDs in the MEC network. Note that, the maximum computation rate need be recomputed by solving (P2) once a WD is turned off/on.
#
#  References:
#  [1] 1. Liang Huang, Suzhi Bi, and Ying-jun Angela Zhang, “Deep Reinforcement Learning for Online Ofﬂoading in Wireless Powered Mobile-Edge Computing Networks”, submitted to IEEE Journal on Selected Areas in Communications.
#
# version 1.0 -- April 2019. Written by Liang Huang (lianghuang AT zjut.edu.cn)
#  #################################################################import scipy.io as sio                     # import scipy.io for .mat file I/
import numpy as np                         # import numpyfrom memory import MemoryDNN
from optimization import bisection
from main import plot_rate, save_to_txtimport timedef WD_off(channel, N_active, N):# turn off one WDif N_active > 5: # current we support half of WDs are offN_active = N_active - 1# set the (N-active-1)th channel to close to 0# since all channels in each time frame are randomly generated, we turn of the WD with greatest indexchannel[:,N_active] = channel[:, N_active] / 1000000 # a programming trick,such that we can recover its channel gain once the WD is turned on again.print("    The %dth WD is turned on."%(N_active +1))# update the expected maximum computation raterate = sio.loadmat('./data/data_%d' %N_active)['output_obj']return channel, rate, N_activedef WD_on(channel, N_active, N):# turn on one WDif N_active < N:N_active = N_active + 1# recover (N_active-1)th channel channel[:,N_active-1] = channel[:, N_active-1] * 1000000 print("    The %dth WD is turned on."%(N_active))# update the expected maximum computation  raterate = sio.loadmat('./data/data_%d' %N_active)['output_obj']        return channel, rate, N_activeif __name__ == "__main__":''' This demo evaluate DROO for MEC networks where WDs can be occasionally turned off/on. After DROO converges, we randomly turn off on one WD at each time frame 6,000, 6,500, 7,000, and 7,500, and then turn them on at time frames 8,000, 8,500, and 9,000. At time frame 9,500 , we randomly turn off two WDs, resulting an MEC network with 8 acitve WDs.'''N = 10                     # number of usersN_active = N               # number of effective usersN_off = 0                  # number of off-usersn = 10000                     # number of time frames, <= 10,000K = N                   # initialize K = Ndecoder_mode = 'OP'    # the quantization mode could be 'OP' (Order-preserving) or 'KNN'Memory = 1024          # capacity of memory structureDelta = 32             # Update interval for adaptive Kprint('#user = %d, #channel=%d, K=%d, decoder = %s, Memory = %d, Delta = %d'%(N,n,K,decoder_mode, Memory, Delta))# Load datachannel = sio.loadmat('./data/data_%d' %N)['input_h']rate = sio.loadmat('./data/data_%d' %N)['output_obj']# increase h to close to 1 for better training; it is a trick widely adopted in deep learningchannel = channel * 1000000channel_bak = channel.copy()# generate the train and test data sample index# data are splitted as 80:20# training data are randomly sampled with duplication if n > total data sizesplit_idx = int(.8* len(channel))num_test = min(len(channel) - split_idx, n - int(.8* n)) # training data sizemem = MemoryDNN(net = [N, 120, 80, N],learning_rate = 0.01,training_interval=10, batch_size=128, memory_size=Memory)start_time=time.time()rate_his = []rate_his_ratio = []mode_his = []k_idx_his = []K_his = []h = channel[0,:]for i in range(n):# for dynamic number of WDsif i ==0.6*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_off(channel, N_active, N)if i ==0.65*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_off(channel, N_active, N)if i ==0.7*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_off(channel, N_active, N)if i ==0.75*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_off(channel, N_active, N)if i ==0.8*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_on(channel, N_active, N)if i ==0.85*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_on(channel, N_active, N)if i ==0.9*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_on(channel, N_active, N)channel, rate, N_active = WD_on(channel, N_active, N)if i == 0.95*n:print("At time frame %d:"%(i))channel, rate, N_active = WD_off(channel, N_active, N)channel, rate, N_active = WD_off(channel, N_active, N)if i % (n//10) == 0:print("%0.1f"%(i/n))if i> 0 and i % Delta == 0:# index counts from 0if Delta > 1:max_k = max(k_idx_his[-Delta:-1]) +1; else:max_k = k_idx_his[-1] +1; K = min(max_k +1, N)i_idx = ih = channel[i_idx,:]# the action selection must be either 'OP' or 'KNN'm_list = mem.decode(h, K, decoder_mode)r_list = []for m in m_list:# only acitve users are used to compute the rater_list.append(bisection(h[0:N_active]/1000000, m[0:N_active])[0])# memorize the largest rewardrate_his.append(np.max(r_list))rate_his_ratio.append(rate_his[-1] / rate[i_idx][0])# record the index of largest rewardk_idx_his.append(np.argmax(r_list))# record K in case of adaptive KK_his.append(K)# save the mode with largest rewardmode_his.append(m_list[np.argmax(r_list)])
#        if i <0.6*n:# encode the mode with largest rewardmem.encode(h, m_list[np.argmax(r_list)])total_time=time.time()-start_timemem.plot_cost()plot_rate(rate_his_ratio)print("Averaged normalized computation rate:", sum(rate_his_ratio[-num_test: -1])/num_test)print('Total time consumed:%s'%total_time)print('Average time per channel:%s'%(total_time/n))# save data into txtsave_to_txt(k_idx_his, "k_idx_his.txt")save_to_txt(K_his, "K_his.txt")save_to_txt(mem.cost_his, "cost_his.txt")save_to_txt(rate_his_ratio, "rate_his_ratio.txt")save_to_txt(mode_his, "mode_his.txt")

DROO demo_on_off.py相关推荐

DROO demo_alternate_weights.py
为方便自己看代码所以放上来,看完就删.完整代码在文章作者的仓库:https://github.com/revenol/DROO. 这个是论文<Deep Reinforcement Learnin ...
DROO memory.py
这个是论文<Deep Reinforcement Learning for Online Ofﬂoading in Wireless Powered Mobile-Edge Computing ...
DROO main.py
是论文<Deep Reinforcement Learning for Online Ofﬂoading in Wireless Powered Mobile-Edge Computing Ne ...
DROO optimization.py
为方便自己看代码所以放上来,看完就删.完整代码在文章作者的仓库:https://github.com/revenol/DROO. 这个是论文<Deep Reinforcement Learnin ...
汉字的首拼音字母生成
生成助记码(取汉字的第一个字母) SET NOCOUNT ON GO IF EXISTS(SELECT name FROM sysobjects WHERE name = N'hzp ...
python零基础对main.py的学习（DROO）
在Python中处理Matlab数据时,需要用到sicpy.io**;** sicpy.io提供两个函数: 1.loadmat: loadmat(file_name[, mdict, appendma ...
DROO源码及论文学习
目录源码描述之optimization.py 公式描述 local计算: 卸载到边: CD法交替方向乘子法ADMM 算法流程 DROO的网络模型略谈边缘计算卸载模型卸载模式信道模型计算模型 ...
解决 win10 pycurl安装出错 Command python setup.py egg_info failed with error code 10 编译安装包安装万金油...
版权声明:本文为博主原创文章,遵循 CC 4.0 BY-SA 版权协议,转载请附上原文出处链接和本声明. 本文链接:https://blog.csdn.net/yexiaohhjk/article/d ...
SyntaxError: Non-ASCII character ‘\xe5‘ in file(xxlrt_1.py) on line 7, but no encoding declared；
具体报错如下: SyntaxError: Non-ASCII character '\xe5' in file /home/hly/work/python/work/xlrt_1.py on line ...

DROO demo_on_off.py

DROO demo_on_off.py相关推荐

最新文章

热门文章