Introduction
In the mid-1980s, software engineering faced a significant "complexity gap." While progress had been made in "programming-in-the-small"—the automatic generation of code for specific algorithms—the "programming-in-the-large" phase remained a manual bottleneck. This phase involves conceptual analysis and architectural design, breaking down complex systems into modular components.
The ESAP (Expert System for Automatic Programming) project, developed at the Milan Polytechnic, was a pioneering attempt to bridge this gap. Unlike contemporary systems that focused on algorithm generation, ESAP was designed to assist human designers in the initial, often informal, phases of problem analysis and structural decomposition.
System Architecture and Methodology
ESAP is structured as a collection of specialized "experts" that collaborate to refine a high-level specification into a formal modular architecture.
Specialized Experts
- Specification Expert: This component provides a "User Representation Language" for incremental problem definition. It allows for informal descriptions that are gradually refined into formal logic.
- Programming In-The-Large Expert: The core of the system, this expert analyzes problem specifications and suggests decompositions into subproblems using a library of design rules.
- Lower-Level Actors Expert: This module determines if a sub-module is simple enough to be handled by automated synthesizers (like the Bidirectional Synthesizer) or retrieved from a library of reusable components.
Inference Mechanism
The system employs a State-Object-Action (S-O-A) strategy. The design process is modeled as a state-space search where a State represents a specific set of modules, an Object is a module requiring further expansion, and an Action is a decomposition rule applied to refine the architecture. Metarules are utilized to prioritize specific design patterns, mimicking the heuristic strategies of human experts to navigate the search space efficiently.
Experimental Evaluation: Banking Domain
To test ESAP's capabilities, the authors applied it to the domain of Bank Current Accounts Management. This domain provided a realistic testbed for "programming-in-the-large" due to its inherent complexity, involving transaction handling, interest computation, and registry updates.
The system demonstrated significant flexibility, often suggesting multiple alternative architectures for the same problem. For instance, it could propose different strategies for interest calculation based on available rules, allowing the human architect to select the most appropriate path.
Key Insights and Results
The implementation of ESAP in Franz-Lisp on a DEC VAX11/780 yielded several key insights:
- Incremental Refinement: ESAP successfully handled incomplete specifications, providing guidance even when initial details were missing.
- Reuse over Generation: The system prioritized the reuse of existing methodologies and domain concepts, reflecting a more practical approach to software production than purely generative models.
- Functional Binding: The system could automatically identify when a module reached a state of "functional binding," meaning it was ready for implementation without further decomposition.
Limitations and Future Outlook
Despite its successes, ESAP faced challenges common to early expert systems. Its knowledge base was heavily domain-specific, requiring manual engineering to adapt to new fields. Furthermore, the system lacked autonomous learning capabilities, relying entirely on pre-programmed rules.
The authors concluded that the future of such systems lies in the automatic acquisition of domain knowledge and a deeper understanding of the qualitative "solving power" of lower-level synthesizers. ESAP remains a foundational example of how AI can support the high-level architectural phases of the software lifecycle.
Summary updated