APS

Table of contents

• General info
• Requirements
• Additional Program Requirement
• How to Use

General info

The program APS is the Reinforcement Learning based Model for Gantt Planning.

Time window scanner is used for solving scheduling problems.

Requirements

This module requires the following modules:

• python=3.5(3.5.6)
• tensorflow-gpu==1.14.0 (install gpu version of tensorflow module)
• tensorflow==1.14.0 (install cpu version of tensorflow module)
• scipy==1.2.1
• pygame
• moviepy
• numpy==1.18.5
• pandas
• matplotlib
• PyCharm under version of 2020.1

Additional Program Requirement

In order to generate gif file, ImageMagik program is also required.

ImageMagik

If you are using window OS ,

you should do some additional works follows...

In config_defaults.py , which has a directory :

C:\Users\user\Anaconda3\envs\'virtual_env_name'\Lib\site-packages\moviepy , change the code

IMAGEMAGICK_BINARY = os.getenv('IMAGEMAGICK_BINARY', 'auto-detect')

into

IMAGEMAGICK_BINARY = os.getenv('IMAGEMAGICK_BINARY', 'C:\Program Files\ImageMagick-7.0.9-Q16\magick.exe')

ImageMagick-7.0.9-Q16 is the ImageMagik version you had installed.

---

How to Use

Setting target date or projects

You can either import schedule by target date or target projects

• Setting target date

  If you want to import training/test data by target date, you can simply send target_start_date and target_finish_date to work.py > import_schedule_by_target_date(filepath, target_start_date, target_finish_date)

   import_schedule_by_target_date('../environment/data/191227_납기일 추가.xlsx', '20190201', '20190601')

• Setting target projects

  If you want to import training/test data by target projects, you can simply send target_projects to work.py > import_schedule_by_target_projects(filepath, target_projects)

      import_schedule_by_target_projects('../environment/data/191227_납기일 추가.xlsx', [3095, 'R873'])

---

• Learning Parameters Configurations

  • Time Window Size

    In train.py > main function

    window = (10, 40)

  ​ You can determine the size of Time Window which scans the total imported schedule

  • Learning Rate

    In train.py > main function

    trainer = tf.train.AdamOptimizer(learning_rate=1e-5)

    A recommanded figure for learning rate is 1e-5

  • Discount Rate

    In train.py > main function

    gamma = .99 # discount rate for advantage estimation and reward discounting

    A recommanded figure for discount rate is 0.99

  • Frequency of target model update(N-step bootstrapping)

    In A3C Algorithm, each of Workers collects samples with its own environments. After a certain number of time-steps , target network(global network) is updated with that samples.

    In train.py > Worker>work

    if len(episode_buffer) == 30 and d != True and episode_step_count != max_episode_length-1: # in this case, frequency of target model update is 30 time-steps

  • Number of Threads

    You can also set up number of threads by changing the number of Workers

    In train.py > main function

    num_workers = nultiprocessing.cpu.count() # Set workers to number of available CPU threads
    if num_workers > 8:
        num_workers = 8 
    workers = []