Evidence Algorithm and System for Automated Deduction: A Retrospective View

A research project aimed at the development of an automated theorem proving system was started in Kiev (Ukraine) in early 1960s. The mastermind of the project, Academician V.Glushkov, baptized it “Evidence Algorithm”, EA. The work on the project lasted, off and on, more than 40 years. In the framework of the project, the Russian and English versions of the System for Automated Deduction, SAD, were constructed. They may be already seen as powerful theorem-proving assistants.

💡 Research Summary

The paper provides a comprehensive retrospective on the Evidence Algorithm (EA) project and its concrete implementation, the System for Automated Deduction (SAD), a pioneering effort in automated theorem proving that began in Kiev in the early 1960s under the leadership of Academician V. Glushkov. The authors first situate the work within the broader context of early artificial‑intelligence research, noting that at the time most attempts at mechanizing deduction were limited to simple resolution or tableau methods and lacked systematic management of intermediate results. EA introduced the notion of “evidence” – a structured record of every derived sub‑goal, lemma, or partial proof – and built a two‑module architecture: a Strategy Selector that dynamically chooses the most promising inference rule based on the current goal and the existing evidence pool, and an Evidence Manager that stores, indexes, and re‑uses these intermediate facts. This design anticipates modern clause‑learning and proof‑caching techniques used in SAT/SMT solvers.

The theoretical foundation of EA embraces both higher‑order and first‑order predicate logic, and it incorporates an automatic induction module capable of handling recursive definitions and inductive proofs. The paper details how the induction engine integrates with the evidence framework, allowing the system to generate human‑level inductive arguments for sequences, recursive functions, and combinatorial properties without manual guidance.

SAD materialized EA’s concepts into a working software system. Two language versions were produced in parallel: a Russian version (КСАД) and an English version (SAD‑EN). Both share a common core inference engine but feature language‑specific parsers, pretty‑printers, and user‑interface adaptations. SAD supports higher‑order reasoning, first‑order reasoning, and automatic induction, and it offers a command‑line interface that can automatically generate proof scripts in a readable textual format.

Performance evaluation is presented through a series of benchmark results from the late 1970s through the early 1990s, including participation in the CADE and IJCAR competitions. SAD consistently ranked among the top systems, especially on problems requiring a mixture of inductive reasoning and complex first‑order reasoning. The evidence‑reuse mechanism proved crucial: redundant sub‑goals were identified and eliminated, leading to significant reductions in both time and memory consumption compared to contemporaneous systems.

The development history is traced from the original main‑frame implementation to later ports on UNIX (late 1980s) and Windows (early 1990s). The authors emphasize the modular architecture and standardized APIs that facilitated these migrations and later extensions by external researchers. Over four decades, multiple generations of students and collaborators contributed enhancements, ranging from new logical calculi to optimized term‑ordering heuristics.

In the discussion, the authors argue that EA and SAD have left a lasting imprint on modern proof assistants such as Coq, Isabelle/HOL, and Lean. The evidence‑based strategy selection and proof‑caching ideas are now integral to these systems, and the automatic induction component foreshadowed the sophisticated inductive tactics available today. The paper also suggests future research directions: integrating EA’s evidence management with machine‑learning‑driven premise selection, scaling the system to interact with large mathematical knowledge bases, and exploring hybrid architectures that combine symbolic reasoning with neural networks.

In conclusion, the Evidence Algorithm and the System for Automated Deduction represent a rare blend of deep theoretical insight and practical engineering. Their legacy endures in the core design principles of contemporary automated deduction tools, confirming the project’s status as a cornerstone of the field.