/DeepQLearning_logistic_solving

Primary LanguagePythonMIT LicenseMIT

Solving logistics with DeepQLearning

TL;DR

After about 100 days/games it achive maximum rewards
eg: Trashing no burgers, making no customers wait

The Game

Rendered game

Basics :

  • There is a random number of customers coming to a burger shop each 10 minutes
  • Each one order randomly one burger

There is more customers if :

  • It's lunch or diner time
  • It's a hollyday
  • It's sunny
  • There is low trafic

Other rules :

  • During lunch they tend to order more bigmac
  • On hollydays they tend to order more specials

Burger rules:

  • The burger shop can make between 3 to 30 burgers each 10 minutes
  • Only one kind is made each time
  • Burgers are trashed after 30 mintes on the shelf

how to win maximum reward ?

  • There should be no customer waiting more than 10 minutes to get his burger
  • There should be no trashed burgers

Result

Ploted rewards

the agent manage to make perfect jujement after 100 days/games using 3 convolution network and 2 hidden layers to make the decision

Install env :

$ conda create --name mcdo --file requirements_conda.txt
$ conda activate mcdo

Try the agent :

$ python ntest.py

Try to beat the agent knowing the rules :

$ python htest.py

See how random choices does :

$ python rtest.py