pycsp3-explain 0.1.2

Explanation tools for PyCSP3 constraint models

PyCSP3-Explain

Explanation tools for PyCSP3 constraint models. Find minimal unsatisfiable subsets (MUS), maximal satisfiable subsets (MSS), and minimal correction sets (MCS) to debug and understand infeasible constraint problems.

Features

• MUS (Minimal Unsatisfiable Subset): Find the minimal set of conflicting constraints
  • mus() - Assumption-based using ACE core extraction
  • mus_naive() - Deletion-based (works with any solver)
  • quickxplain_naive() - Preferred MUS based on constraint ordering
• Optimal MUS (SMUS/OCUS): Prefer smaller or lower-weight explanations
  • smus() - Smallest MUS (fewest constraints)
  • optimal_mus() - Minimum total weight MUS
  • ocus() - OCUS with subset constraints/predicates
  • ocus_naive() - Naive OCUS (alias for weighted MUS)

  Note: IHS uses a small CP model in PyCSP3 and falls back to enumeration if needed.

• MSS (Maximal Satisfiable Subset): Find the maximum satisfiable portion
  • mss() - Assumption-based with core extraction
  • mss_naive() - Greedy growing algorithm
• Weighted MSS/MCS: Optimize which constraints to keep/remove
  • mss_opt() - Maximize total weight of kept constraints
  • mcs_opt() - Minimize total weight of removed constraints
• MCS (Minimal Correction Set): Find the minimal changes needed to restore satisfiability
  • mcs() - Via assumption-based MSS complement
  • mcs_naive() - Via naive MSS complement
• MARCO: Enumerate all MUSes and MCSes
  • marco() - Generator over MUS/MCS
  • marco_naive() - Naive MARCO implementation
  • all_mus() / all_mcs() - Convenience wrappers
  • all_mus_naive() - Naive MUS enumeration

Installation

# Clone the repository
git clone https://github.com/sohaibafifi/pycsp3-explain.git
cd pycsp3-explain

# Install dependencies
pip install -e .

Requirements

• Python 3.10+
• PyCSP3 (>= 2.5)
• lxml (>= 4.9)
• Java runtime (for ACE/Choco)

Quick Start

Using satisfy() with Explain Tools

from pycsp3 import *
from pycsp3_explain import *


x = VarArray(size=2, dom=range(10))

satisfy( [
    x[0] == 5,
    x[0] == 7,  # conflicts with x[0] == 5
    x[1] >= 3,
])

constraints  = flatten_constraints(posted())
if solve() == UNSAT:
    print("MUS:", mus(constraints))
    print("Optimal MUS:", optimal_mus(constraints, weights=[1, 1, 1]))
    for kind, subset in marco(constraints):
        print(kind, subset)

    # or use the helper function
    mus_result = explain_unsat("mus", soft=constraints, check=False)
    print("MUS:", mus_result)

    optimal = explain_unsat("optimal_mus", soft=constraints, check=False, weights=[1] * len(constraints))
    print("Optimal MUS:", optimal)

    for kind, subset in explain_unsat("marco", soft=constraints, check=False):
        print(kind, subset)

Notes

• mus() and mss() use ACE core extraction for efficiency; they fall back to naive versions for non-ACE solvers.
• MCS is the complement of MSS: MCS = Soft \ MSS
• Removing MCS constraints from soft constraints leaves MSS, which is satisfiable.

How It Works

Constraint Explanation Problem

Given an infeasible constraint satisfaction problem (CSP), explanation tools help identify:

Concept	Question	Description
MUS	Why is it infeasible?	Minimal Unsatisfiable Subset - the smallest set of constraints that conflict
MSS	What can be satisfied?	Maximal Satisfiable Subset - the largest set of constraints that can hold together
MCS	What to fix?	Minimal Correction Set - the smallest set of constraints to remove for satisfiability

Relationships

• Every MUS and every MCS share at least one constraint (hitting set duality)
• MCS = Soft \ MSS (MCS is the complement of MSS)
• Multiple MUSes/MSSes/MCSes may exist for a given problem

API Reference

MUS Functions

• mus(soft, hard=None, solver="ace") - Assumption-based MUS
• mus_naive(soft, hard=None, solver="ace") - Deletion-based MUS
• quickxplain_naive(soft, hard=None, solver="ace") - Preferred MUS
• is_mus(subset, hard=None, solver="ace") - Verify MUS validity
• all_mus_naive(soft, hard=None, solver="ace") - Enumerate MUSes (naive)
• optimal_mus(soft, hard=None, weights=None, solver="ace") - Minimum-weight MUS
• optimal_mus_naive(soft, hard=None, weights=None, solver="ace") - Naive minimum-weight MUS
• smus(soft, hard=None, solver="ace") - Smallest MUS (fewest constraints)
• ocus(soft, hard=None, weights=None, solver="ace", ...) - OCUS with subset constraints
• ocus_naive(soft, hard=None, weights=None, solver="ace") - Naive OCUS
• OCUSException - Exception for optimal MUS/OCUS routines

MSS/MCS Functions

• mss(soft, hard=None, solver="ace") - Assumption-based MSS
• mss_naive(soft, hard=None, solver="ace") - Greedy growing MSS
• mss_opt(soft, hard=None, weights=None, solver="ace") - Weighted MSS
• is_mss(subset, soft, hard=None, solver="ace") - Verify MSS validity
• mcs(soft, hard=None, solver="ace") - Assumption-based MCS
• mcs_naive(soft, hard=None, solver="ace") - Naive MCS
• mcs_opt(soft, hard=None, weights=None, solver="ace") - Weighted MCS
• mcs_from_mss(mss, soft) - Complement of MSS
• is_mcs(subset, soft, hard=None, solver="ace") - Verify MCS validity

MARCO Enumeration

• marco(soft, hard=None, solver="ace") - Generator over MUS/MCS
• marco_naive(soft, hard=None, solver="ace") - Naive MARCO
• all_mus(soft, hard=None, solver="ace") - Collect all MUSes
• all_mcs(soft, hard=None, solver="ace") - Collect all MCSes

License

MIT License

References

• PyCSP3 - Python library for constraint programming
• CPMpy - Constraint Programming and Modeling in Python
• Liffiton et al. (2016) - "Fast, flexible MUS enumeration"
• Junker (2004) - "QuickXplain: Preferred explanations and relaxations"