Qualitative Reasoning With Higher-Order Derivatives

Johan de Kleer, Daniel G. BobrowAAAI 1984

Abstract

This paper explores the extension of qualitative physics through the introduction of higher-order derivatives into qualitative calculus. By formalizing constraints based on the Mean Value Theorem and Taylor expansions, the authors enable AI systems to predict complex long-term behaviors, such as oscillation and asymptotic limits, without requiring precise numerical data.

Introduction

In the early development of Artificial Intelligence, qualitative reasoning (QR) aimed to help computers understand the physical world as humans do: through causal relationships rather than complex differential equations. While researchers successfully modeled immediate changes, they struggled with long-term "gross-time" behaviors. The primary obstacle was ambiguity; for example, a first-order model cannot determine if a decreasing positive value will eventually reach zero or stabilize. This paper by Johan de Kleer and Daniel Bobrow (1984) addresses this by formalizing how higher-order derivatives (e.g., acceleration and jerk) constrain the possible future states of a system.

Qualitative Calculus and the Six Laws

The authors propose a "qualitative calculus" that operates on signs (+, 0, -). Their core contribution is a set of six fundamental laws used to prune impossible transitions in a state-transition graph:

  1. Value Continuity: Values must change smoothly (e.g., + to - requires passing through 0).
  2. Contradiction Avoidance: Prunes states that violate the system's governing qualitative equations.
  3. Instant Change Rule: Changes from zero happen before changes to zero.
  4. Derivative Continuity: Continuity rules applied to derivatives.
  5. Higher-Order Derivatives: Uses Taylor expansions to determine which derivative "wins" in determining the next state.
  6. No Change to All Zero Derivatives: A non-zero quantity cannot become identically zero at an instant.

Benchmark Systems

The theory is evaluated using two classic physical benchmarks: the pressure regulator and the mass-spring-friction system. The latter is particularly significant for demonstrating the modeling of oscillation.

The Mass-Spring-Friction system used as a benchmark for qualitative oscillation modeling.

In a mass-spring system, first-order reasoning fails to explain why a mass overshoots its equilibrium point. Higher-order derivatives provide the necessary "momentum" to describe this behavior qualitatively.

Key Results: Modeling Oscillation

The algorithm performs a "generate and test" simulation to identify physically reachable states. In the mass-spring-friction example, the algorithm initially identifies 15 possible transitions from a given state. By applying the six laws, it prunes 12 of them as physically impossible.

Comparison between (a) a quantitative damped sine wave and (b) the qualitative state-transition cycle generated by the algorithm.

As shown in Figure 4, the resulting state-transition graph (b) mirrors the qualitative phases of a damped sine wave (a). The system identifies a cycle of states (detailed in the paper's tables as 13 distinct states) that capture the essence of decaying oscillation without numerical computation.

Limitations and Future Work

A significant challenge remains in modeling "quiescence"—the point where a system stops. While common sense suggests a friction-based system eventually rests, the qualitative math (following Newtonian laws) suggests an asymptotic approach to zero. The authors suggest that "qualitative quantum mechanics" or specific axioms may be needed to handle these limits. Additionally, as the order of a system increases, the potential for state-space explosion and ambiguity in higher-order signs remains a topic for future research.

Conclusion

By incorporating higher-order derivatives, de Kleer and Bobrow significantly advanced the ability of AI to reason about physical systems. This work proves that symbolic AI can capture the fundamental behavior of complex calculus, providing a bridge between simple causal logic and rigorous mathematical modeling.

Summary updated

BibTeX
@inproceedings{kleer1984qualitative,
  author = {Johan de Kleer and Daniel G. Bobrow},
  title = {{Qualitative Reasoning With Higher-Order Derivatives}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1984},
}
Figure