Neuro-Symbolic Reasoning: Fundamentals, Models, Certification, and Systems

Contents

Neuro-Symbolic Reasoning: Fundamentals, Models, Certification, and Systems#

— Knowledge graph–driven neuro-symbolic reasoning and trustworthy governance in low-altitude traffic: theory and practice#

Scope of the book#

Core objective: Using “knowledge graph–driven neuro-symbolic reasoning and trustworthy governance” as the main line, this book explains the full technical chain of neuro-symbolic AI—from foundational theory and key models through certification mechanisms to system deployment—and takes urban low-altitude traffic / UAM as the primary scenario to develop a unified methodology that is interpretable, calibratable, deployable, and governable.

This book includes theoretical material related to the open-source project liuyushugreat/SkyNetUamPlatform.

Introduction: From the two-peaks dilemma in AI to the third path of neuro-symbolic AI#

Core objectives#

Clarify the historical tension between symbolic and connectionist AI; explain why neuro-symbolic AI has again become central in safety-critical complex systems; and introduce trustworthy governance of low-altitude traffic as the overarching problem of the book.

Chapter 1: The evolution, fracture, and reconstruction of AI#

  • 1.1 The rise and fall of symbolic AI: from expert systems to the knowledge-engineering bottleneck

  • 1.2 The rise of connectionism: deep learning’s success and the black-box crisis

  • 1.3 The two-peaks dilemma: the long-standing split between strong perception and weak reasoning

  • 1.4 Lessons from cognitive science: complementary mechanisms of System 1 and System 2

  • 1.5 The proposal of neuro-symbolic AI: why unify perception, knowledge, reasoning, and decision-making

  • 1.6 Special requirements in safety-critical settings: explanation, certification, robustness, real time

  • 1.7 Low-altitude traffic / UAM: a representative challenge for neuro-symbolic AI

  • 1.8 Main line and technical closed loop of this book: knowledge base—reasoning models—trustworthy certification—system deployment

Part I: Foundations—logic, representation, and learning#

Core objectives#

Equip readers from diverse backgrounds with foundations in logic, knowledge representation, deep learning, and graph learning, laying a shared base for later neuro-symbolic integration.

Chapter 2: Symbolic logic, rule systems, and automated reasoning#

  • 2.1 Propositional logic and first-order predicate logic

  • 2.2 Rules, constraints, and reasoning chains

  • 2.3 Forward chaining, backward chaining, and resolution

  • 2.4 Description logic and ontology languages: foundations

  • 2.5 Uncertain logic and fuzzy logic: an introduction

  • 2.6 From formal logic to engineering rule systems

Chapter 3: Knowledge graphs and domain knowledge representation#

  • 3.1 Basic concepts of knowledge graphs: entities, relations, attributes, events

  • 3.2 RDF, OWL, SPARQL, and graph query languages

  • 3.3 Ontology construction: concept layer, relation layer, constraint layer

  • 3.4 Event graphs, scene graphs, and temporal knowledge graphs

  • 3.5 From multi-source heterogeneous data to unified knowledge representation

  • 3.6 Knowledge graph updating, alignment, and consistency maintenance

  • 3.7 From general KGs to domain-specific KGs

Chapter 4: Deep learning, graph learning, and neural representation#

  • 4.1 Basic ideas of deep representation learning

  • 4.2 Inductive biases of neural networks and function approximation

  • 4.3 Graph neural networks: message passing, relational modeling, graph representation

  • 4.4 Transformer and attention mechanisms

  • 4.5 Temporal modeling and learning dynamical systems

  • 4.6 Limits of deep learning: OOD behavior, hallucination, catastrophic forgetting, fragility

  • 4.7 From deep models to constrained learning

Part II: A unified knowledge base—domain modeling for low-altitude traffic#

Core objectives#

Make explicit a pivotal chapter derived from the author’s dissertation work (Chapter 3): build the shared knowledge base for all subsequent models.

Chapter 5: Domain knowledge modeling for trustworthy governance in low-altitude traffic#

  • 5.1 Object system for low-altitude traffic / UAM: aircraft, airspace, missions, infrastructure, stakeholders

  • 5.2 Modeling risk objects: conflict, violation, failure, environmental disturbance

  • 5.3 Modeling rule objects: regulations, operational constraints, priority rules, safety boundaries

  • 5.4 Spatiotemporal objects: routes, corridors, windows, dynamic occupancy

  • 5.5 Scene semantic templates: logistics, inspection, emergency, air taxi

  • 5.6 Risk labels and design of explanation units

  • 5.7 From business language to knowledge language: unified semantic abstraction

Chapter 6: Low-altitude traffic knowledge graph construction and unified representation#

  • 6.1 Ontology layering: concept layer, instance layer, rule layer

  • 6.2 Entity–relation patterns and attribute modeling

  • 6.3 Unified expression of static knowledge, dynamic state, and certification objects

  • 6.4 Encoding rule constraints: logical rules, soft constraints, conflict resolution

  • 6.5 Graph construction pipeline: acquisition, cleaning, mapping, fusion, updating

  • 6.6 Multi-source data ingestion: telemetry, maps, weather, policy, and operation logs

  • 6.7 How the unified knowledge base supports subsequent models

Chapter 7: Taxonomy and technology roadmap for neuro-symbolic systems#

  • 7.1 Classical taxonomy: Henry Kautz’s neuro-symbolic spectrum

  • 7.2 A three-layer integration taxonomy: symbolic-dominant / hybrid-pipeline / deeply fused

  • 7.3 Layer 1: rule-dominant knowledge-augmented systems

  • 7.4 Layer 2: hybrid pipelines with KG + retrieval augmentation + language models

  • 7.5 Layer 3: deeply fused models running directly on KG / TKG

  • 7.6 Comparing the three paths: interpretability, real time, scalability, certification readiness

  • 7.7 Main technical line of the book and chapter mapping

Part III: Static cognition—knowledge graph–driven interpretable risk reasoning#

Core objectives#

Explain how to build static interpretable reasoning frameworks with KG + RAG + LLM in settings that are rule-heavy, knowledge-dense, and demand very high interpretability.

Chapter 8: Basic paradigms of knowledge injection and constrained learning#

  • 8.1 Loss-function formulations of logical rules

  • 8.2 Constrained optimization and posterior regularization

  • 8.3 Co-training of rules and neural models

  • 8.4 Physical and mechanistic constraint networks

  • 8.5 From knowledge injection to interpretable reasoning

Chapter 9: Hybrid neuro-symbolic reasoning driven by knowledge graphs#

  • 9.1 SPARQL queries and deterministic evidence extraction

  • 9.2 Semantic completion by LLMs and chain-style explanation generation

  • 9.3 Applicability and limits of RAG in rule-dense scenarios

  • 9.4 GraphRAG: from document retrieval to graph retrieval

  • 9.5 Coupling external solvers with language models

  • 9.6 From “being able to answer” to “answering with evidence”

Chapter 10: Designing interpretable risk reasoning frameworks—from SkyKG to a general methodology#

  • 10.1 Problem definition for static risk reasoning

  • 10.2 Coordinated organization of ontology + rules + retrieval indexes

  • 10.3 Dual-channel reasoning: rule-query channel and semantic generation channel

  • 10.4 Risk label outputs and explanation-chain generation

  • 10.5 Faithfulness evaluation: rule alignment, unsupported-claim rate, evidence coverage

  • 10.6 Strengths and boundaries of static single-body scenarios

  • 10.7 Why move from static cognition to dynamic coordination

Part IV: Dynamic coordination—temporal knowledge graphs and multi-agent real-time reasoning#

Core objectives#

Focus on high-density, strongly time-varying, multi-agent interaction; explain how temporal KGs and graph neural networks support real-time conflict detection and coordinated de-confliction.

Chapter 11: Temporal knowledge graphs and dynamic relational reasoning#

  • 11.1 From static KG to temporal KG

  • 11.2 Representing time, events, and state transitions

  • 11.3 Dynamic relations and time-varying edge weights

  • 11.4 Temporal logic and continuous-time modeling

  • 11.5 Event-driven reasoning and dynamic updates

  • 11.6 Applicability of dynamic graphs in traffic and flight systems

Chapter 12: Graph-driven multi-agent conflict detection models#

  • 12.1 Formalizing multi-UAV dynamic coordination

  • 12.2 Graph attention and relation-aware message passing

  • 12.3 Temporal encoding and stale-edge discounting

  • 12.4 Spatiotemporally coupled conflict prediction

  • 12.5 Real-time metrics: latency, throughput, stability

  • 12.6 Accuracy metrics: CDR, FAR, F1, and warning lead time

  • 12.7 From prediction to intervention: dynamic risk scores and priority ranking

Chapter 13: Coordinated de-confliction and decision-making on temporal relational graphs#

  • 13.1 Relationship between conflict detection and conflict resolution

  • 13.2 Generating avoidance strategies from paths and graph structure

  • 13.3 Role of reinforcement learning in graph-path reasoning

  • 13.4 Constraint-based coordinated de-confliction

  • 13.5 Local optimality vs. global safety envelopes

  • 13.6 From SkyFlow to a general dynamic reasoning framework

Part V: Trustworthy certification—calibratable, provable, auditable neuro-symbolic reasoning#

Core objectives#

Move “trustworthy AI” from loose discussion toward certification-grade trustworthy reasoning—the most dissertation-distinctive part of the book.

Chapter 14: Trustworthiness issues in neuro-symbolic systems#

  • 14.1 Why neuro-symbolic systems are still not inherently trustworthy

  • 14.2 Interpretability ≠ certifiability

  • 14.3 Black-box confidence and overconfidence

  • 14.4 Distribution shift, concept drift, and scenario transfer

  • 14.5 Special trustworthiness requirements in safety-critical systems

  • 14.6 From empirical performance to formal assurance

Chapter 15: Conformal prediction and uncertainty calibration#

  • 15.1 Types of uncertainty: epistemic vs. aleatoric

  • 15.2 Why probabilistic outputs are not the same as trustworthy confidence

  • 15.3 Basic principles of conformal prediction

  • 15.4 Conformal methods in classification, regression, and set prediction

  • 15.5 Conformal calibration for graph and temporal models

  • 15.6 Conformal scoring design for neuro-symbolic outputs

  • 15.7 Risk prediction intervals and coverage guarantees

Chapter 16: Online monitoring, distribution shift, and statistical certification#

  • 16.1 Failure modes in online inference

  • 16.2 Martingale methods and sequential consistency monitoring

  • 16.3 Concept drift and anomalous-distribution detection

  • 16.4 Trust scoring for static explanation chains

  • 16.5 Certification wrapping for dynamic conflict prediction

  • 16.6 From “interpretable output” to “certification-grade output”

  • 16.7 SkyCert: systematic design of a certification bridge

Chapter 17: Interpretability evaluation, audit chains, and regulatory interfaces#

  • 17.1 Layers of interpretability: local, rule-based, causal

  • 17.2 Evaluating faithfulness, sufficiency, and comprehensibility

  • 17.3 Audit log design for neuro-symbolic systems

  • 17.4 How human supervisors use explanations and certificates

  • 17.5 Interface issues with aviation / autonomous-driving safety standards

  • 17.6 From model trustworthiness to governance trustworthiness

Part VI: System foundations—city-scale cloud–edge neuro-symbolic reasoning architecture#

Core objectives#

Corresponding to SkyGrid: land all prior algorithms in city-scale systems engineering.

Chapter 18: Complexity of neuro-symbolic reasoning and the compute gap#

  • 18.1 Sources of complexity in symbolic, graph, and large-model reasoning

  • 18.2 Latency constraints of real-time systems

  • 18.3 Concurrency pressure in city-scale low-altitude traffic

  • 18.4 From algorithmic optimality to deployable utility

  • 18.5 Necessity of layered deployment of the reasoning stack

Chapter 19: Distributed reasoning systems under cloud–edge collaboration#

  • 19.1 Division of roles between edge and cloud reasoning

  • 19.2 Distributed orchestration of rule checking, GNN inference, and LLM inference

  • 19.3 Streaming data ingestion and event-driven scheduling

  • 19.4 Knowledge sync, model sync, and state consistency

  • 19.5 Real-time alerting chains and closed-loop response

  • 19.6 From single-node feasibility to multi-node scale-out

Chapter 20: Spatiotemporal graph partitioning and high-concurrency reasoning engines#

  • 20.1 Spatiotemporal graph partitioning

  • 20.2 Subgraph cuts, boundary coordination, and cross-region conflict propagation

  • 20.3 Parallelism mechanisms in graph reasoning engines

  • 20.4 Load balancing for large-scale conflict detection

  • 20.5 Throughput and fault tolerance for city-scale thousand-UAV scenarios

  • 20.6 SkyGrid: from research prototype to engineering base

Chapter 21: Platform realization—from stacking models to governance closed loops#

  • 21.1 Integrating knowledge base—reasoning models—certification—deployment architecture

  • 21.2 Modular design for platforms

  • 21.3 Unifying data flow, control flow, and evidence flow

  • 21.4 Multi-role interfaces for regulators, operators, and executors

  • 21.5 From research systems to industry application platforms

  • 21.6 Open-source ecosystems and reproducible experiments

Part VII: Frontiers—large models, agents, and AI for Science#

Core objectives#

Keep the textbook forward-looking without leaving the main line—extensions beyond the core thread.

Chapter 22: Neuro-symbolic AI in the large language model era#

  • 22.1 LLMs as System 1: strengths, hallucination, structural limitations

  • 22.2 Coupling language models with symbolic rules

  • 22.3 GraphRAG and knowledge-augmented reasoning

  • 22.4 LLMs calling external solvers and toolchains

  • 22.5 Neuro-symbolic agents: perceive—retrieve—reason—act

  • 22.6 Agent coordination for complex governance systems

Chapter 23: Safety-critical industry applications#

  • 23.1 Intelligent transportation and autonomous driving: scene-graph understanding

  • 23.2 Urban low-altitude traffic: risk assessment and dynamic coordination

  • 23.3 Interpretable reasoning stacks in medical diagnosis

  • 23.4 Industrial control and fault diagnosis in complex systems

  • 23.5 From “usable” to “deployable with confidence”: key differences

Chapter 24: Neuro-symbolic reasoning in AI for Science#

  • 24.1 Molecular graphs, scientific knowledge, and interpretable discovery

  • 24.2 Physical constraints and scientific mechanism learning

  • 24.3 Knowledge-augmented simulation in complex systems

  • 24.4 Prospects for neuro-symbolic methods in cross-disciplinary science

  • 24.5 Feedback from industry applications to foundational theory

Closing: Summary and future research roadmap#

Chapter 25: The future of neuro-symbolic reasoning—from interpretability toward certification, governance, and deployment#

  • 25.1 Summary of the main line: knowledge base, reasoning, certification, systems

  • 25.2 Current bottlenecks: data, standards, benchmarks, certification, compute

  • 25.3 Future direction I: unified world models and neuro-symbolic agents

  • 25.4 Future direction II: certification-grade trustworthy AI

  • 25.5 Future direction III: real-time governance of city-scale complex systems

  • 25.6 Future direction IV: from UAM to broader safety-critical systems

Appendices and labs#

Appendix A: Mathematical, logical, and graph-learning notation#

  • A.1 Logical symbols and common rule templates

  • A.2 Notation for graph neural networks and temporal graph models

  • A.3 Conformal prediction and statistical calibration notation

  • A.4 System architecture and complexity-analysis notation

Appendix B: Common open-source libraries and engineering tools#

  • B.1 PyKEEN

  • B.2 DeepProbLog

  • B.3 Logic Tensor Networks (LTN)

  • B.4 PyTorch Geometric (PyG)

  • B.5 RDFLib / OWLReady2 / Neo4j / GraphDB

  • B.6 LLM + KG integration toolchains

  • B.7 Streaming inference and edge deployment frameworks

Appendix C: Course labs and practice projects#

  • Lab 1: Small low-altitude-traffic ontology construction and rule encoding

  • Lab 2: Static risk reasoning with KG + SPARQL + LLM

  • Lab 3: Conflict detection models on temporal knowledge graphs

  • Lab 4: Conformal prediction for risk outputs with confidence intervals

  • Lab 5: Prototype cloud–edge collaborative reasoning system

  • Lab 6: Joint agent development with large models + knowledge graph + solvers

For the English introduction and reading guide, see INTRODUCTION_EN.md.

Contents