Conflict Driven Clause Learning

University of Washington | CSE 442

We have a sudoku board!

You may assume a solution exists, but how can we check?

The Boolean Satisfiability Problem

The Boolean Satisfiability Problem is one in which we have to decide if there exists an assignment for some boolean variables that would make a boolean formula evaluate to true.

If a formula has an assignment under which the formula evaluates to True, then it is satisfiable. If no such assignment exists for a formula, then it is unsatisfiable.

In general, the Boolean Satisfiability Problem can take a very long time to solve (in fact, it’s NP-Complete). If a boolean formula has n variables, then there are 2^n possible assignments. All known algorithms, in the worst case, will have to do a full search on that exponentially sized search space. However, in real world problems, there are often logical structures in the problem that an algorithm can utilize to search better.

In this article, we will be covering two algorithms that try to find a satisfying assignment for a formula (SAT solvers). The first one will be the Davis-Putnam-Logemann-Loveland Algorithm developed in the 1960s, which forms the basis for modern SAT solving algorithms today. The second one will be the Conflict-Driven Clause Learning Algorithm which is more recent - around 1996, which improves on the first algorithm in really cool ways.

Both algorithms will only work on boolean formulas that are in Conjunctive Normal Form (CNF). In this form, boolean formulas are composed of the ANDs (∧) of clauses, clauses are the ORs (v) of literals, and literals are a variable or its negation.

Formulas in CNF have some nice properties that make reasoning quick. Since the formula is just the clauses ANDed together, for our formula to be satisfied every clause must be satisfied. Since each clause is made of the ORs of literals, only one literal needs to be satisfied for the whole clause to be satisfied.

Below is the example formula f = (-x1 ∨ -x2 ∨ x3) ∧ (-x3 ∨ x4). See if you can find assignments that satisfy this formula.

The visualization starts out by visualizing the formula in a single line. This can be unwieldy for large formulas. We will visualize it instead by having each clause on its own line. Clicking the “To Clauses” and the “To Formula” button will show you the difference in the visualizations. Just remember that all clauses must be true for the formula to be true.

Boolean Constraint Propagation

When an algorithm searches for a satisfying assignment for a CNF formula, one optimization is to look for unit clauses. A clause is unit under a partial assignment when that assignment makes every literal in the clause unsatisfied but leaves a single literal undecided.

Because the clause must be satisfied for the formula to be satisfied, there’s no point in checking any assignment that would make that undecided literal false. The algorithm can just make its next guess such that the literal will be true and thus the clause true.

This process is known as Boolean Constraint Propagation (BCP). Algorithms will run BCP to assign variables that logically only have one option due to unit clauses. Sometimes, doing BCP for unit clauses will cause other clauses to become unit! Algorithms will run BCP repeatedly until there are no unit clauses.

In the example below, we have guessed so far that x1 and x2 are both true. Try to predict how BCP will work below.

Davis-Putnam-Logemann-Loveland Algorithm

The DPLL algorithm is a SAT solver based on recursive backtracking that makes use of BCP. While a naive recursive backtracker would check every possible assignment until it found one that satisfies the formula, DPLL makes BCP-driven inferences to cut down on the size of the search tree – removing assignments that can’t satisfy the formula. When it encounters a conflict, it backtracks to the last non-BCP decision it made, and makes the other choice.

Here’s the pseudo-code:

DPLL:
  Run BCP on the formula.
  If the formula evaluates to True, return True.
  If the formula evaluates to False, return False.
  If the formula is still Undecided:
    Choose the next unassigned variable.
    Return (DPLL with that variable True) || (DPLL with that variable False)

See how DPLL works with the example below. On the left, we have a visualization of the decision tree which shows the history of choices that the algorithm makes. Edges that are solid represent the guesses/decisions the algorithm makes, while edges that are dashed represent that the algorithm used BCP to get that choice. The path in blue leads to where the algorithm currently is at.

Since there is nothing to BCP, DPLL starts with some guessing.

Conflict Driven Clause Learning

DPLL has three shortcomings. First, it makes naive decisions. Second, when it encounters a conflict, it doesn’t learn anything else from the conflict besides the fact that the current partial assignment led to the conflict. And third, it can only backtrack one level which can cause it to keep exploring a search space that is doomed to fail.

Conflict Driven Clause Learning (CDCL) improves on all three aspects! We’ll focus on the improvements that CDCL makes on the last two shortcomings: clause learning from conflicts, and non-chronological backtracking.

Without going into detail, let’s take a look at CDCL in action on the same example!

CDCL behaves the same as DPLL for the first 3 steps.

Awesome! It found that the formula was satisfiable much faster. Let’s take a closer look at how CDCL operates.

Clause Learning

When CDCL reaches a conflict, it looks at the guesses it has made, and all of the assignments it was forced to infer via BCP which eventually led to the conflict. If we graph out these guesses and their implications, we get a graph called the implication graph. This is the graph of the decisions it made, the literals that were BCP’d and why, and how it got to the conflict. By looking at this graph, CDCL is able to learn a clause that is potentially more useful than just the knowledge that the current partial assignment was bad. This way, CDCL can avoid making the same mistake over and over and skip over large chunks of bad partial assignments DPLL will get stuck in.

What’s great about the learned clause is that it allows CDCL to learn from its mistakes and ignore huge sections of the search space that will never satisfy the formula.

Once CDCL has learned a clause, CDCL will be able to backtrack potentially more than one level, and BCP using the learned clause to put what it learned into action immediately. The ability to backtrack more than one level with the learned clause is what we call non-chronological backtracking.

Watch how CDCL constructs the implication graph below.

Since there is nothing to BCP, CDLL can start with some guessing.

A note about Decision Heuristics

The version of CDCL we implemented in this article does not have any smart decision heuristics. Better decision heuristics is often the source of breakthroughs SAT Solver performance increases and are very important to modern SAT solvers today. We only examined clause learning and non-chronological backtracking as improvements over DPLL to keep the scope of this project small.

Solving Sudoku

Sudoku is what we started this article with so lets try to solve it with our SAT solvers!

To solve Sudoku, we encode the four rules into CNF clauses.

  1. Cells must be filled in with a number between 1 and 9.
  2. Rows can not have duplicate numbers.
  3. Columns can not have duplicate numbers.
  4. Subsquares can not have duplicate numbers.

The encoding we use is to make 9 propositional variables per cell, each representing a possible number for that cell, and then writing out the rules among those variables.

When we get partial assignments back from the solver, we decode it back to sudoku board assignments and then show them.

To wrap up, take a look below to see how DPLL does vs. CDCL! Notice how much less manual decisions there are for CDCL. That is because it is learning about the logical structure of sudoku through its mistakes - super cool.

DPLL

# of manual assignments: 0

CDCL

# of manual assignments: 0

References