Running solver

To run the solver, simply use:

vrp-cli solve pragmatic problem.json -o solution.json -g solution.geojson --log

If you specify --log option, it will produce some log output which contains various information regarding refinement process such as costs, amount of routes, time, etc.:

configured to use max-time: 300s
configured to use custom heuristic
preparing initial solution(-s)
[0s] created initial solution in 536ms, fitness: (0.000, 104.000, 70928.735)
[1s] created initial solution in 513ms, fitness: (0.000, 109.000, 74500.133)
[2s] created initial solution in 1097ms, fitness: (0.000, 104.000, 70928.735)
[2s] created initial solution in 168ms, fitness: (0.000, 125.000, 94305.015)
created initial population in 2317ms
[2s] generation 0 took 21ms, fitness: (0.000, 104.000, 70669.056)
[2s] population state (phase: initial, speed: 0.00 gen/sec, improvement ratio: 1.000:1.000):
        rank: 0, fitness: (0.000, 104.000, 70669.056), difference: 0.000%
        rank: 1, fitness: (0.000, 104.000, 70705.550), difference: 0.052%
[5s] generation 100 took 27ms, fitness: (0.000, 96.000, 64007.851)
[7s] generation 200 took 19ms, fitness: (0.000, 95.000, 63087.282)
..
149s] generation 4000 took 44ms, fitness: (0.000, 92.000, 54032.930)
[149s] population state (phase: exploration, speed: 26.78 gen/sec, improvement ratio: 0.235:0.155):
        rank: 0, fitness: (0.000, 92.000, 54032.930), difference: 0.000%
        rank: 1, fitness: (0.000, 92.000, 54032.930), difference: 0.000%
[153s] generation 4100 took 42ms, fitness: (0.000, 92.000, 54021.021)
..
[297s] generation 7200 took 20ms, fitness: (0.000, 92.000, 53264.644)
[299s] population state (phase: exploitation, speed: 24.16 gen/sec, improvement ratio: 0.165:0.058):
        rank: 0, fitness: (0.000, 92.000, 53264.026), difference: 0.000%
        rank: 1, fitness: (0.000, 92.000, 53264.026), difference: 0.000%
[299s] total generations: 7246, speed: 24.16 gen/sec
Route 1: 144 925 689 739 358 32 783 924 461 111 766 842 433
..
Route 92: 837 539 628 847 740 585 328 666 785 745
Cost 53264.03

Once the problem is solved, it will save solution in pragmatic and geojson (optional) format.

Extra options

The vrp-cli supports extra command line arguments which affects behavior of the algorithm.

Search mode

By default, search is mostly performed in exploration mode (broad) which allows the algorithm to perform better exploration of a solution space. However, it has slower overall convergence in better local optimum. You can switch to exploitation mode with deep setting:

vrp-cli solve pragmatic problem.json --search-mode=deep

In this mode, the algorithm memorizes only the last discovered best known solutions, so it can jump quicker to relatively good local optimum, but suffers more from premature convergence.

A general recommendation is to use deep on relatively simple dataset and/or when strict time limits should be applied.

Heuristic mode

At the moment, the solver supports three types of hyper-heuristics:

• static selective: chooses metaheuristic from the list of predefined within their probabilities
• dynamic selective: applies reinforcement learning technics to adjust probabilities of predefined metaheuristics
• multi selective (default): starts with dynamic selective and switches to static selective if the progression speed is slow

You can switch between modes with heuristic setting:

vrp-cli solve pragmatic problem.json --heuristic=static

Termination criteria

Termination criteria defines when refinement algorithm should stop and return best known solution. At the moment, there are three types which can be used simultaneously:

Max time

Max time specifies duration of solving in seconds:

vrp-cli solve pragmatic problem.json --max-time=600

Max generations

Generation is one refinement step and it can be limited via max-generations parameter:

vrp-cli solve pragmatic problem.json --max-generations=1000

Coefficient of variation

This criteria calculates coefficient of variation for each objective over specific amount of generations specified by sample and stops algorithm when all calculated values are below specified threshold. It can be defined by min-cv parameter:

vrp-cli solve pragmatic problem.json --min-cv=sample,200,0.1,true

Here, the first parameter is generation amount, the second - threshold, the third - boolean flag whether the logic is applicable for all search phases, not only exploitation.

Alternatively, period of time in seconds can be specified instead of sample:

vrp-cli solve pragmatic problem.json --min-cv=period,300,0.1,true

Due to internal search heuristic implementation, it is recommended to use this termination criteria with max-time or max-generations.

Default behavior

Default termination criteria is max 3000 generations and 300 seconds at max.

Initial solution

You can supply initial solution to start with using -i option. Amount of initial solutions to be built can be overridden using init-size option.

Writing solution to file

Writing solution into file is controlled by -o or --out-result setting. When it is omitted, then solution is written in std out.

Geojson

Pragmatic format supports option -g or --geo-json which writes solution in separate file in geojson format. If the library iw used from the interop api (e.g. python or c), then solution in geojson can be returned inside extras.features with the config option specified: s

{
  "output": {
    "includeGeojson": true
  }
}

s