Real-world AI problems rarely fit neatly into a single mathematical box. A logistics plan requires both discrete logic—"if the truck is in Chicago, it cannot be in New York"—and continuous numerical reasoning—"the truck has 50 gallons of fuel and consumes 5 gallons per hour." Historically, researchers solved these problems by picking a side: either discretizing the numbers into logical symbols or treating the logic as a set of rigid equations.
However, a growing body of research suggests that the most efficient way to solve these hybrid problems is not to choose between logic and math, but to unify them into a single, tightly coupled engine.
The Cost of Separation
When logical reasoning and numerical solving are treated as isolated modules, the system suffers from a "combinatorial explosion." In a decoupled architecture, a Boolean Satisfiability (SAT) solver might suggest thousands of logically valid assignments that are physically impossible due to resource or time constraints. The numerical solver only sees these failures after the logic solver has already done the work.
Beringer et al. argue that this modular decomposition leads to significant information loss. By integrating logical backtracking with numerical techniques like bound propagation and linear relaxation, the authors demonstrate that a unified solver can prune the search space much earlier. The intuition is simple: the solver should be able to "see" a numerical inconsistency as soon as a partial logical assignment makes it inevitable, rather than waiting for a complete solution.
Tight Coupling and Interleaved Search
The core mechanism of this unified approach is the interleaved search. Instead of solving the logic first, the engine updates numerical constraints dynamically as logical assignments are made.
One of the earliest successful implementations of this was the LPSAT Engine, which integrated a SAT solver with a Simplex-based linear programming solver. This allowed the system to handle "Linear Constraint Normal Form" (LCNF), where propositional variables and linear inequalities are solved in tandem. Similarly, Shin et al. showed that this integration is essential for planning in continuous-time domains, where autonomous processes (like a tank filling with water) must be balanced against discrete events.
Learning from Mathematical Conflict
A major breakthrough in this field involves how solvers learn from their mistakes. In pure logic, when a solver hits a dead end, it generates a "nogood"—a rule that says "this specific combination of choices is impossible."
Drescher et al. extended this to numerical constraints through "lazy nogood generation." When a mathematical constraint (like a resource limit) is violated, the system translates that numerical failure back into a logical rule. This allows the logic solver to apply conflict-driven learning to mathematical constraints, effectively "learning" the boundaries of the physical world as it searches.
Bridging High-Level Logic and Low-Level Solvers
This research direction has evolved into sophisticated frameworks like Constraint Answer Set Programming (CASP). As Formisano et al. note, the goal is to allow high-level declarative languages to leverage optimized low-level search engines.
This synergy is particularly powerful in complex scheduling. For example:
- Bao et al. used a combination of Answer Set Programming (ASP) and Constraint Logic Programming to handle "temporally expressive" planning without the need to discretize time into tiny, inefficient chunks.
- Geibinger et al. applied CASP to industrial laboratory scheduling, using logic to handle complex dependencies and numerical solvers to manage precise timing.
- Eiter et al. demonstrated that by separating job assignment logic from temporal arithmetic using specialized difference logic propagators, solvers could find optimal schedules for semiconductor manufacturing that were previously unreachable.
By treating logic and math as two sides of the same coin, these unified solvers avoid the bottlenecks of traditional methods, enabling AI to reason about the world with both the precision of a mathematician and the rigor of a logician.
Key Foundational Work
The LPSAT Engine & Its Application to Resource Planning (1999) — contributed 2 concept elements to this direction.
Satisfiability of Boolean Formulas over Linear Constraints (1993) — contributed 2 concept elements to this direction.
Satisfiability of Boolean Formulas over Linear Constraints (1993) — contributed 1 concept elements to this direction.