HotNets'16 http://people.csail.mit.edu/hongzi/content/publications/DeepRM-HotNets16.pdf
Install prerequisites
Python 2.7
sudo apt-get update
sudo apt-get install python-numpy python-scipy python-dev python-pip python-nose g++ libopenblas-dev git
pip install --user Theano==0.8
pip install --user Lasagne==0.1
sudo apt-get install python-matplotlib
In folder RL, create a data/ folder.
Set environment variable MKL\_THREADING\_LAYER=GNU
Use launcher.py to launch experiments.
- cuDNN v4
- Edit theano config in launcher.py
--exp_type <type of experiment>
--num_res <number of resources>
--num_nw <number of visible new work>
--simu_len <simulation length>
--num_ex <number of examples>
--num_seq_per_batch <rough number of samples in one batch update>
--eps_max_len <episode maximum length (terminated at the end)>
--num_epochs <number of epoch to do the training>
--time_horizon <time step into future, screen height>
--res_slot <total number of resource slots, screen width>
--max_job_len <maximum new job length>
--max_job_size <maximum new job resource request>
--new_job_rate <new job arrival rate>
--dist <discount factor>
--lr_rate <learning rate>
--ba_size <batch size>
--pg_re <parameter file for pg network>
--v_re <parameter file for v network>
--q_re <parameter file for q network>
--out_freq <network output frequency>
--ofile <output file name>
--log <log file name>
--render <plot dynamics>
--unseen <generate unseen example>
The default variables are defined in parameters.py.
Example:
- launch supervised learning for policy estimation
python launcher.py --exp_type=pg_su --simu_len=50 --num_ex=1000 --ofile=data/pg_su --out_freq=10
- launch policy gradient using network parameter just obtained
python launcher.py --exp_type=pg_re --pg_re=data/pg_su_net_file_20.pkl --simu_len=50 --num_ex=10 --ofile=data/pg_re
- launch testing and comparing experiemnt on unseen examples with pg agent just trained
python launcher.py --exp_type=test --simu_len=50 --num_ex=10 --pg_re=data/pg_re_1600.pkl --unseen=True