Introduction
In classical artificial intelligence, a persistent challenge has been the trade-off between computational efficiency and system robustness. Associational rule-based systems, while fast, are often brittle when encountering scenarios outside their predefined "if-then" patterns. Conversely, causal models that reason from first principles are robust but computationally expensive, often leading to an exponential search space. This paper introduces the Generate, Test and Debug (GTD) paradigm, implemented in the GORDIUS system, which seeks to marry the speed of compiled knowledge with the deep reasoning of causal simulations.
The GTD Paradigm
The core of the GTD approach is a three-stage cycle that moves beyond simple success/failure feedback. Instead of restarting the search upon failure, the system attempts to repair its initial hypothesis using causal insights.
- Generate: The system employs "scenarios"—associational rules that map observable topological patterns in data back to likely causes. This stage provides a fast, "mostly right" initial hypothesis.
- Test: A simulator, termed the "imaginer," performs qualitative and quantitative simulations of the hypothesis. If a discrepancy is found between the simulated result and the goal, the tester generates a causal explanation—a dependency tree identifying the specific assumptions that led to the failure.
- Debug: This stage is the primary innovation. The debugger uses the causal explanation to pinpoint the "bug" and applies domain-independent strategies (e.g., inserting a process to change a state) alongside domain-specific models to repair the plan.
Geologic Interpretation and Domain Testing
The primary evaluation domain for GORDIUS was geologic interpretation: inferring the sequence of events (such as deposition, erosion, and faulting) that produced a specific vertical cross-section of the Earth.
In this domain, the generator might suggest a sequence that includes erosion. However, if the tester determines the land was below sea level at that time, the debugger identifies this as a violation of the preconditions for erosion. It then consults its causal models to find a process that increases height—such as "uplift"—and inserts it into the sequence to resolve the conflict. Beyond geology, the GTD paradigm was successfully tested in Blocks World planning, the Tower of Hanoi puzzle, and semiconductor manufacturing diagnosis.
Key Insights and Results
- Debugging as Search Guidance: By transforming a failure into a rich causal map, GTD drastically narrows the search space compared to traditional generate-and-test methods.
- Knowledge Compilation: Associational rules are treated as "compiled" versions of causal knowledge. They provide efficiency, while the underlying causal model remains available as a fallback for complex reasoning.
- Robustness through Hybridization: The system successfully handles "unforeseen interactions" where the independence assumptions of the generator break down. The hybrid approach proved more robust than rules alone and more efficient than pure first-principles planning.
Conclusion and Future Work
While GORDIUS demonstrates the power of the GTD paradigm, its effectiveness is inherently tied to the accuracy of its underlying causal models. Future research is needed to determine the optimal "switching point"—the moment when a hypothesis is so fundamentally flawed that it is more efficient to regenerate it from scratch rather than continue debugging. Despite these challenges, GTD represents a significant step toward AI systems that can reason both quickly and deeply.
Summary updated