This is a Automatic parameter tunning program for Minizinc solvers.

Currently support cplex, gurobi and osicbc and able to perform parameter tunning on multiple instances. The program use SMAC v2.10.03 to perform parameter tunning.

Cplex: Tested on cplex 12.8

Gruobi: Tested on gurobi 8.1. Having issues when using gurobi 7.5, cannot reproduce tuning result.

1. Java (Tested on 8, not sure does it working on 9 or 10)

   SMAC2 is written in java.

2. Minizinc

1. Download the latest released version and unzip it.

2. Run the "setup" program to let Minizinc and Minizinc IDE known where is the tuning program

3. You can add the tuning program folder to enviroment path if you use command line version frequently.

1. On default, the program will try to minimize the time cost ot obtain optimal solution and use the parameter space configuration files in pcsFiles/ folder.

2. Go to the directory of your model files and data files

3. Use a command in following format to start tunning:

Path/to/Minizinc-Tunning --solver [1.solver] -p [2.No.of threads] -t [3.time limit] [4.model and data] 

[1]: Solver you want to use: cplex / gurobi / osicbc

[2]: How many threads that minizinc use to solve the problem.

Attention the prameter configration for certain amount processors may not work when using another processors setting. This means if you use -p 2 for parameter tunning, the tunning result may not improve the model solving speed if you run minizinc with -p 1. (-p is not necessary. If remove -p, it will use default value 1) You also can tune No. of threads, see details in Advanced part.

[3]: Total time limit for tunning the program, unit is seconds.

The time limit should be enough for runing the model for at leat 80 times. If your models needs 300 seconds to solve the problem, it is suggested to tuning for more than 6 hours (21600 seconds). If you only tunning for 3 hours, it will also give you a improved parameter configuration, but may not as good as 6 hour result. It is suggested to make it run as long as possible.

[4]: models and datas: "Model1.mzn data1.dzn data2.dzn" "Model2.mzn data21.dzn" "Model3.mzn data31.dzn data32.dzn data33.dzn ..." ....

[5]: If you are using cplex on linux， or minizinc don't known where is cplex, add --cplex-dll [5.dll file path] to the end of command. [5.dll file path] is the path of cplex dll/so file. Remove it if you are using osicbc or gurobi

../Minizinc-Tunning --solver cplex -p 2 -t 3600 "mapping.mzn mesh3x3_mp3_2.dzn ring_mp3.dzn" 

--obj-mode

If your model is hard to achieve a optimal solution in reasonable time or your want to improve the optimal solution. You can use this argument to optimize the objective of solution within limited time. Remember to use -c [cut off time] to set time limit

On default it try to minimize objective

For maximization problem, must add

--maximize

to indeticate that this is a maximization problem

and must add

-c to specify a cutoff time.

-psmac [No. of smac]

SMAC support parallel search and tunning. SMACs will share their data to biuld model. If your computer have enough CPU resource, use this to achieve better solution. For cpu with 8 threads, if you run minizinc with -p 1, you can use -psmac 6. If minizinc use -p 2 , you can try -psmac 3. Do not use all aviable threads, as cpus are not only used by minizinc and smac.

--tune-threads

With this argument, the program will treat threads as one of parameter for tunning. -p [No. of threads] will be treat as maximum threats available and default threats setting.

-pcsJson

If you have your own parameter space configuration file, you can use this argument to provide the path of it.

-c

Set cut off time for each minizinc run manualy

-v

Enable verbose/debug mode

--skip-bench

Skip final benchmark, use smac-output to judge and output best configuration.