Abstract
The Modular Action Description (MAD) language is introduced to address the lack of generality in AI reasoning about actions. By extending the action language $C+$ with modularity, MAD enables the creation of reusable knowledge libraries. This paper details the language's syntax, its formal translation into causal logic, and its application to classic AI benchmarks like the Monkey and Bananas domain.
Introduction
In artificial intelligence research, "toy worlds" such as the Blocks World or the Missionaries and Cannibals puzzle serve as standard environments for testing reasoning about change. However, a persistent inefficiency exists: researchers must redefine the fundamental "physics" of these worlds for every new model. For instance, "climbing" a box and "walking" across a room are both specialized forms of "moving," yet they are often coded as entirely separate rules. This lack of a shared, reusable knowledge base is the "key problem of generality in AI" identified by John McCarthy. MAD addresses this by providing a framework for modularity—building complex systems from simple, swappable, and reusable parts.
Methods: Extending Causal Logic
MAD is built as an extension of the action language $C+$, which is adept at describing causal relationships but lacks hierarchical organization. MAD introduces several key features:
- Module Imports: Modules can inherit fluents (changing properties) and actions from other modules.
- Renaming Clauses: Generic terms in imported modules can be specialized. For example, a generic "Thing" in a
MOVEmodule can be renamed to "Monkey" for a specific scenario. - Formal Translation: The language employs a rigorous mathematical process to "flatten" modular descriptions into standard $C+$ descriptions, ensuring compatibility with existing logical solvers.
- Causal Logic Foundation: By utilizing nonmonotonic causal logic, MAD handles complex issues like the "frame problem" (indirect effects) and concurrent actions more effectively than older systems like STRIPS.
Key Insights: Action Inheritance and Elaboration Tolerance
A central innovation of MAD is treating actions similarly to objects in object-oriented programming. Just as a "Square" inherits from a "Shape," a "Climb" action in MAD inherits the logic of a general "Move" action while adding specific constraints, such as vertical movement. This approach enhances elaboration tolerance, allowing researchers to modify specific behaviors in one module without rewriting foundational logic. This mirrors human language, where specific verbs like "walk" are defined as special cases of more general actions like "move."
Results and Evaluation
The efficacy of MAD is demonstrated through its application to classic AI benchmarks, most notably the Monkey and Bananas domain. The language allows for a concise description of this complex scenario by importing and specializing a generic MOVE module.
Crucially, the authors show that MAD's modular structures can be mapped onto transition diagrams—directed graphs where vertices represent states and edges represent actions. This provides a clear mathematical path for AI solvers to process modular descriptions. The paper also establishes a "standard model" for MAD, ensuring consistent logical conclusions across different implementations.
Limitations and Future Work
While MAD provides the necessary language, the actual "general-purpose database of actions" remains to be constructed. Future efforts will focus on integrating MAD directly into tools like the Causal Calculator (CCALC), which currently requires manual translation from MAD to $C+$. Additionally, the authors aim to extend the language to handle "non-definite" causal laws, where actions may have multiple possible outcomes.
Summary updated