incaseoftrouble / stationary-distribution-sampling

Code for the associated paper

Stationary Distribution Sampling

Building

• Extract sources if necessary
• Run docker build -t sds .

Running

• To run models: Execute docker run --rm --cpus=1 sds timeout 10m time -v ... with

  • PRISM: prism -steadystate <model path> -const <model constants>
  • Classic: sds <model path> --const <model constants> solve
  • Naive: sds <model path> --const <model constants> approximate --precision 1e-4 --sampling SAMPLE_NAIVE
  • Sample: sds <model path> --const <model constants> approximate --precision 1e-4 --sampling SAMPLE_TARGET

• Models from the paper:

  • brp: models/dtmc/brp/brp.pm, N=64,MAX=5
  • nand: models/dtmc/nand/nand.pm, N=15,K=2
  • zeroconf_dl: models/mdp/zeroconf_dl/zeroconf_dl.nm, reset=false,deadline=40,N=1000,K=1
  • phil4: models/mdp/phil/phil4.nm (no constants) (add --explore to the SDS call)
  • rabin3: models/mdp/rabin/rabin3.nm (no constants) (add --explore to the SDS call)
  • branch: models/dtmc/crafted_branch/branch.pm, loops=1000,steps=100,p=0.008,levels=1000

• Some further interesting models:

  • brp: models/dtmc/brp/brp.pm, N=100,MAX=100
  • crowds: models/dtmc/crowds/crowds.pm, TotalRuns=3,CrowdSize=10
  • nand: models/dtmc/nand/nand.pm, N=10,K=5
  • loop: models/dtmc/crafted_loop/loop.pm, size=1000 (add -explicit to PRISM call)
  • error: models/dtmc/crafted_error/error.pm, p=0.5,e=0.0000001 (add -maxiters 1000000000 to PRISM the call to increase the maximal iterations and -power or -jor to see the wrong results under the respective iteration schemes) Note: PRISM's hybrid engine does eventually converge to the correct result, see below.
  • error_hybrid: models/dtmc/crafted_error/error_hybrid.pm, e=0.0000001 (a variant model where the default hybrid scheme immediately terminates and yields a wrong solution)

• To execute the full benchmark suite run: docker run --cpus=1 -t -v $(pwd)/results:/opt/results:rw sds python3 run.py --results /opt/results/data.json --timelimit 60 <further args> (or --timelimit 300 to fully replicate the experiments in the paper)

• This can also be configured to exclude some approaches or models, --help explains further arguments

• To view results, run docker run -v $(pwd)/results:/opt/results:ro sds python3 run.py /opt/results/data.json). Append --only single_scc or --only one_state_scc to apply the respective filters. Add --scatter approx,solve to obtain the data for scatter plots.