PaperlayerView exact revision on arXiv

Source-linked AI summary

Ontology-Mediated Neurosymbolic Constraint Acquisition from Multiple Stakeholders

Stefan Bischof, Juliana Kainz, Danilo Valerio

arXiv:2609.29876v1cs.AI

TL;DR

Neurosymbolic systems often assume a pre-existing symbolic specification, leaving the acquisition of constraints from multiple sources underexplored. The paper proposes an OWL ontology-mediated loop for integrating stakeholder preferences with hardware limits and returning explanations for renegotiation; its FLEXI example distinguishes individually impossible requests from preferences that are feasible alone but unmet jointly.

  • Problem

    Neurosymbolic work often assumes a given specification, while real constraint sets may need reconciliation across heterogeneous sources with conflicting requirements and unequal authority.

    Source: §1 · p. 1

  • Method

    An OWL configuration ontology integrates LLM-elicited soft preferences and hardware hard limits, checks conflicts, and returns symbolic explanations for assistant-led renegotiation.

    Source: §1 · p. 2, §3.1 · p. 4, §3.2 · p. 5

  • Results

    In FLEXI, an individually impossible charging request is flagged before scheduling, while individually feasible requests may still be jointly unmet when many vehicles charge at once.

    Source: §4 · p. 6

  • Takeaways & Limitations

    The architecture is argued to apply beyond energy where constraints come from multiple human and automated sources with unequal authority.

    Source: §5 · p. 6

    The paper treats the scheduler as a downstream consumer and leaves making that scheduler neurosymbolic outside its scope.

    Source: §4 · p. 5

Abstract

from arXiv · show

Neurosymbolic research typically assumes a pre-existing symbolic specification, leaving the upstream challenge of acquiring and formalizing requirements and constraints largely unaddressed. We present an architecture that fills this gap by using an OWL configuration ontology to mediate between neural constraint sources and downstream consumers. In this framework, LLM assistants elicit soft stakeholder preferences, while hardware specifications define hard physical and engineering limits. The ontology unifies these heterogeneous inputs, leverages description logic to identify unsatisfiability, and generates symbolic explanations that enable LLMs to interactively renegotiate terms with users. Any remaining conflicts are resolved downstream via priority-based relaxation. We illustrate our approach on a microgrid use case from the FLEXI project and argue its generalizability to multi-stakeholder domains where constraint acquisition is distributed across human and automated sources of unequal authority.

1. Introduction

Neurosymbolic systems commonly check neural outputs against existing specifications, but acquiring a consistent specification from multiple sources of unequal authority is less explored. The paper proposes an ontology-mediated architecture for this upstream task and illustrates it in the FLEXI microgrid.

1. Introduction

  • Constraint acquisition remains an underexplored upstream challenge because heterogeneous sources can supply mutually inconsistent requirements with unequal authority.Hardware and safety limits are non-negotiable, while stakeholder preferences may conflict and carry different weights.Source: §1 · p. 1
  • The paper proposes a configuration ontology as a mediating layer, addressing CONTO’s lack of preference, priority, and optimization semantics.CONTO represents components, features, and constraints and supports Description-Logic consistency checking, but treats every constraint as mandatory.Source: §1 · pp. 1–2
  • The architecture integrates LLM assistants and hardware sources into an OWL ontology with relaxable, prioritized preferences and a conflict-explanation renegotiation loop.The paper frames multi-stakeholder acquisition as a complement to downstream validation and instantiates the architecture on a microgrid.Source: §1 · p. 2
  • In FLEXI, EV owners state charging preferences, operators set infrastructure requirements, and hardware defines physical limits.A request to charge 90 kWh into a 60 kWh battery is impossible, while a full charge by 16:00 may become unschedulable when many vehicles charge together.Source: §1 · p. 2

2. Related Work

Related work often constrains neural outputs using an already-given specification or acquires constraints under a single implicit authority. This paper instead focuses on reconciling concurrently acquired constraints from multiple stakeholders with unequal authority.

2. Related Work

  • Constrained decoding and symbolic solvers or provers constrain or verify neural outputs, generally assuming the specification is already given.Logic-LM iterates generation, checking, and revision for one given problem, rather than reconciling competing stakeholder constraints.Source: §2 · p. 2
  • Constraint-acquisition methods learn networks from valid and invalid assignments or user queries, typically assuming one implicit authority.They rarely combine natural-language dialogue, an explicit ontology, and differing stakeholder priorities.Source: §2 · p. 2
  • Recent LLM and ontology or constraint-programming work supports requirements elicitation and configuration, while this paper targets reconciliation across conflicting sources.The sources may have unequal authority and contribute constraints concurrently.Source: §2 · p. 2

3. Approach

The approach places an ontology between constraint sources and a downstream consumer, extending configuration representations to include relaxable preferences. It checks whether preferences are individually compatible with hard limits, returns symbolic explanations for conflicts, and leaves jointly competing preferences to downstream priority-based relaxation.

3. Approach

  • The approach organizes an acquisition loop around an ontology that mediates between constraint sources and a downstream consumer.Constraints are checked as they arrive; the scheduler remains a downstream consumer outside the acquisition loop.Source: §3 · p. 2, §3.1 · p. 3

3.1. Architecture Overview

  • CONTO provides instance-based ConfigVoc and axiom-based ConfigDL representations, but both treat constraints as hard; the preference extension is illustrated on ConfigVoc.How to add preference semantics to either representation remains open, particularly because relaxable constraints do not fit OWL’s all-or-nothing axiom semantics cleanly.Source: §3.1 · p. 3
  • The proposed vocabulary marks constraints hard or soft, assigns priorities or weights to soft constraints, and groups each stakeholder’s preferences into an aggregate.Hard constraints must hold; relaxing a soft constraint incurs a priority-proportional violation cost, keeping preferences in the same representation as physical limits.Source: §3.1 · p. 4

3.2. Consistency Checking and the Feedback Loop

  • The static check tests each soft preference as if it were hard against the hard constraints, detecting structural conflicts through description logic and numeric bound violations through arithmetic reasoning.ConfigDL can use an OWL DL reasoner and SWRL arithmetic rules for grounded values; ConfigVoc can export formula constraints to a solver such as MiniZinc or ASP.Source: §3.2 · pp. 4–5
  • Passing the static check guarantees only individual satisfiability; the scheduler handles jointly infeasible preferences by relaxing them according to priorities and weights while preserving hard constraints.This second relaxation point resolves quantitative and resource feasibility and reports the realized schedule through the assistant.Source: §3.2 · p. 5
  • When a preference clashes with hard limits, the system returns a minimal set of responsible named constraints so the originating assistant can renegotiate it with the stakeholder.The explanation is a symbolic justification rather than free text or a scalar signal, and the loop iterates toward a statically feasible set.Source: §3.2 · p. 5

4. Use Case: The Microgrid

The microgrid example distinguishes preferences that are impossible under hard limits from individually feasible requests that conflict when many vehicles compete for capacity. A downstream scheduler relaxes lower-priority preferences while preserving hard limits.

4. Use Case: The Microgrid

  • The downstream scheduler forecasts variable factors such as photovoltaic production and grid-energy CO2 intensity; making that scheduler neurosymbolic is outside this paper’s scope.Source: §4 · p. 5
  • EV-owner and operator assistants elicit preferences and infrastructure requirements, while hardware specifications supply physical limits.Examples include departure time and desired charge, stationary-battery reserves, fuse and transformer capacities, battery safety parameters, and market obligations.Source: §4 · p. 5
  • A feasible full charge by 16:00 can still become an 80% charge by 16:00 when simultaneous demand exceeds available capacity, so the scheduler relaxes lower-priority preferences while holding hard limits.The shortfall occurs because the transformer limit and reserved stationary-battery power leave too little capacity to meet every target in time.Source: §4 · p. 6
  • A request that exceeds an EV battery’s capacity is impossible under hard limits, and the static check returns a justification so the assistant can ask the owner to revise it before scheduling.The example is 90 kWh requested for a battery that holds 60 kWh.Source: §4 · p. 6

5. Discussion and Conclusion

The paper frames multi-stakeholder constraint acquisition as a neurosymbolic problem and instantiates an ontology-mediated architecture on the FLEXI microgrid. It argues the architecture may apply to other domains, while identifying evaluation and design questions that remain open.

5. Discussion and Conclusion

  • The architecture is instantiated on the FLEXI microgrid, with the hard/soft distinction and explanation-and-renegotiation loop proposed for resolving conflicts across sources of unequal authority.Source: §5 · p. 6
  • Multi-tenant building climate control is presented as a structurally similar non-energy setting, with tenant preferences competing for supply under facility and equipment constraints.Source: §5 · p. 6
  • The work describes an architecture rather than an evaluated system; its preference extension and assistants are not yet built, and the feedback loop has been exercised only on the OWL path.Open questions include priority semantics across stakeholder groups, faithful language-to-ontology grounding, and interactive-time reasoning.Source: §5 · p. 6

Declaration on Generative AI

The authors report using Claude for wording, style, and language checks, then reviewing and editing the content and taking responsibility for the publication.

Declaration on Generative AI

Loading 2609.29876v1…