Architecting Self-Adaptive Systems with Learned and Symbolic Components

AI-enabled self-adaptive systems increasingly combine learned components with symbolic components. Learned components, such as perception, prediction, or language models, handle uncertain, high-dimensional inputs, while symbolic components, such as planners, monitors, policies, and runtime constraints, provide explicit structure and checkable decision logic. However, the architectural boundary at which learned outputs become symbolic facts, constraints, or actions often remains underspecified. Scores, labels, embeddings, or language fragments may cross component boundaries without explicit assumptions about calibration, freshness, ontology, or policy compliance, forcing downstream components to rely on implicit interpretations. This dissertation investigates how such learned-symbolic boundaries can be specified, monitored, and evolved as first-class architectural seams in self-adaptive systems. The proposed abstraction is the seam contract: a connector-level specification consisting of typed boundary objects, versioned constraint bundles, runtime monitors, and decision traces. The research follows a design-science methodology and uses a Goal–Question–Metric evaluation to assess expressiveness, runtime assurance, and bounded revalidation in a simulated perception-to-decision pipeline and a policy-governed LLM assistant. ... mehrAt this stage, the framework and evaluation plan are defined; next steps are refining the contract model, implementing seamKIT and instantiating both scenarios.