Chapter 2: Intelligent systems: A hybrid approach

Introduction

The introduction frames intelligent advisor systems as combining KGs and LLMs to tackle complex problems, and argues that understanding their operation is necessary to judge their trustworthiness and safety.

• Hybrid intelligent system capabilities: Explain that the chapter presents KGs and LLMs together as a way to address highly complex problems by combining existing knowledge and context with reasoning and natural language understanding.
  Why it matters: This is the chapter’s starting point for discussing intelligent advisor systems.
• Trustworthy system assessment: Justify why understanding how intelligent systems operate internally matters when judging their trustworthiness and safety in critical applications.
  Why it matters: A realistic view of a system’s capabilities and limits supports informed decisions about critical uses.

Key terms: knowledge graphs (KGs), large language models (LLMs), intelligent systems, trustworthiness, safety, limitations, genuine capabilities

2.1 What is intelligence?

The section defines intelligence in terms of acquiring and applying knowledge, then introduces knowledge representation and reasoning as the interacting functions of an intelligent system.

• Intelligence as applied knowledge: Explain intelligence as acquiring and applying knowledge to learn from experience, solve problems, and interact with the environment. Describe why designing an intelligent system calls for breaking those activities into components.
  Why it matters: This gives you a way to explain what the system must do before discussing its components.
• Representation and reasoning roles: Distinguish knowledge representation from reasoning and explain how they depend on each other. Describe the trade-off the section identifies when choosing a representation.
  Why it matters: The distinction explains both what an intelligent system works with and what it does with that information.
• Interacting system components: Describe the high-level architecture in which multiple knowledge bases and reasoning engines can coexist and interact. Trace the section's example of one process feeding another.
  Why it matters: The diagram is a flexible mental model, not a fixed two-step pipeline.
• Knowledge graphs with LLMs: Explain the roles this section assigns to knowledge graphs and to their combination with LLMs, without assuming a particular implementation.
  Why it matters: It states why the book brings KGs and LLMs together in an intelligent system.

Key terms: intelligence, intelligent agent, knowledge representation, reasoning, expressive, efficiently, knowledge bases, reasoning engines, knowledge refinement, KGs

2.2 Designing an intelligent system

The section uses a medical diagnosis example to introduce two design ideas: combining general knowledge with patient context, and connecting algorithms so their outputs work together.

• General knowledge and patient context: Explain why the proposed medical diagnostic system combines general knowledge with contextual information when supporting doctors’ choices.
  Why it matters: A diagnostic answer needs both broad medical knowledge and information relevant to the situation at hand.
• Connected algorithm outputs: Describe how algorithms and tools can work together in a system intended to unlock the power of machine learning.
  Why it matters: The section presents an intelligent system as a combination of working parts, rather than a single isolated algorithm.

Key terms: global generic knowledge, contextualized information, system of algorithms and tools, output, input

2.2.1 What is an intelligent system?

Defines an intelligent system as a user-centered application of AI and ML that improves over time through data and user interaction.

• User-centered intelligent systems: Explain that an intelligent system connects users to AI and ML to help them achieve meaningful objectives, especially by improving their decisions rather than replacing them.
  Why it matters: It keeps the system’s purpose focused on the people whose work it supports.
• Improvement through ongoing interaction: Describe how an intelligent system improves over time by learning from user interactions and feedback while using context and maintaining an evolving knowledge base.
  Why it matters: It distinguishes an improving intelligent system from one that merely applies a fixed store of information.

Key terms: intelligent system, users, AI and ML, decision-making capabilities, user interactions, explicit feedback, contextual information, evolving knowledge base, data sources

2.2.2 Categories of intelligent systems

This section distinguishes systems that act for users from systems that advise them, then explains why the book focuses on intelligent advisory systems.

• Autonomous and advisory systems: Distinguish intelligent autonomous systems from intelligent advisory systems by who makes and carries out decisions. Describe the autonomous system’s independence and the advisor’s recommendation-only role.
  Why it matters: The distinction clarifies whether a system is responsible for acting or for helping a person decide.
• Advisory features and KG fit: Explain how decision support, context awareness, and user interaction characterize IASs and why these features make them a focus for using knowledge graphs.
  Why it matters: It explains why the book emphasizes KG-based tools that help people make decisions rather than replace them.

Key terms: intelligent autonomous systems, intelligent advisory systems (IASs), decision support, intelligent advisor systems, context awareness, user interaction, KGs

2.2.3 Characteristics of an intelligent system

The section identifies four characteristics that should guide an intelligent system’s design and applies them to an autonomous advisor that uses domain knowledge and learns from feedback.

• Purpose and intelligent experience: Explain why an intelligent system needs a specific, achievable objective and how its user interface should deliver useful outputs, manage mistakes, and gather feedback.
  Why it matters: The system’s purpose and interface determine whether its intelligence is useful and safe for users.
• Knowledge lifecycle and orchestration: Describe why intelligent behavior requires ongoing knowledge creation and update, and why the system must coordinate multiple algorithms and tools.
  Why it matters: An intelligent system must keep its knowledge useful and coordinate its parts, not rely on a single static model.
• Autonomous advisor design choices: Explain the section’s design choices for an autonomous advisor: suggesting rather than taking actions, grounding advice in expert-refined domain knowledge, and learning from outcomes.
  Why it matters: These choices define what the advisor is allowed to do, what grounds its advice, and how it improves.

Key terms: meaningful objective, intelligent experience, implicit and explicit user feedback, knowledge creation and update, orchestration, LLMs, KGs, autonomous advisor systems, established knowledge base, domain-specific understanding

2.3 Knowledge acquisition and representation

This section explains how intelligent systems acquire and represent knowledge through KGs and LLMs, and why their differences make them complementary.

• Knowledge acquisition in intelligent systems: Explain how an intelligent system acquires knowledge from domain data, user feedback, and its environment, then uses a stored representation during reasoning.
  Why it matters: It connects the information an intelligent system receives to the knowledge it can use for reasoning.
• Explicit knowledge graph representation: Describe how KG acquisition transforms data into nodes and relationships, guided by domain schemas or ontologies, and why this preparation takes care.
  Why it matters: The explicit structure makes a KG interpretable, but building a faithful one requires attention to the domain.
• Implicit language model representation: Describe how LLM training encodes unstructured text as statistical patterns rather than schema-defined relationships, including the benefit and limitation of that representation.
  Why it matters: It explains both the language strengths of LLMs and why their stored knowledge is less directly accessible than a KG’s.
• Practical differences in stored knowledge: Distinguish KG and LLM knowledge by how accessible it is, how it is updated, and what each technology is naturally suited to do.
  Why it matters: These differences help explain what each technology contributes to an intelligent system.
• Complementary hybrid knowledge representations: Justify combining KGs and LLMs using their different representations and the kinds of reasoning each supports.
  Why it matters: The hybrid approach uses the strengths of both representations rather than requiring either one to handle every kind of knowledge alone.

Key terms: knowledge acquisition, knowledge representations, knowledge bases, reasoning phase, KGs, nodes, edges, ontologies, schemas, domain expert

2.4 Reasoning

This section explains how reasoning in an intelligent system draws on supplied information, contrasts deductive and inductive reasoning, and argues for combining KGs and LLMs while recognizing their limits.

• Reasoning inputs and uncertainty: Describe how a user request supplies goals and information to an IAS reasoning engine, which delivers insights and suggestions. Explain why uncertainty in the starting information matters.
  Why it matters: A plausible conclusion is only as dependable as the information it starts from.
• Deduction and inductive generalization: Distinguish deduction from induction by their starting points and conclusions. Use the spam-filter example to explain why matching previously labeled emails is not the same as learning to classify unseen ones.
  Why it matters: The distinction explains both how a system can go beyond stored examples and why doing so introduces uncertainty.
• LLM reasoning on altered problems: Explain what the simplified farmer-and-sheep example illustrates about LLM reasoning, and relate it to the reported results on variants of familiar tasks.
  Why it matters: Fluent, familiar-looking steps do not necessarily show that a model solved the problem actually posed.
• Complementary KG and LLM strengths: Explain which reasoning tasks the section assigns to KGs and to LLMs, and why their combination still does not amount to human common-sense reasoning.
  Why it matters: Choosing a hybrid design requires matching each approach to its strengths without assuming the combination solves every reasoning problem.

Key terms: IAS, reasoning engine, uncertainty, deductive reasoning, inductive reasoning, inductive inference, generalize, LLMs, probabilistic reasoning engine, counterfactual variants

2.5 Reasoning engines

This section explains how a reasoning engine and knowledge base form a feedback loop in an intelligent system, while noting that the engine may use one or several reasoning strategies.

• Reasoning engine feedback loop: Trace how the reasoning engine reads from and writes to the knowledge base, generates actions or suggestions that influence the environment, and processes new observations to build knowledge for subsequent responses.
  Why it matters: The feedback loop explains how an intelligent system can adapt its responses as its environment changes.
• Flexible reasoning strategies: Explain that the reasoning engine shown is generic: it may use one reasoning strategy, such as deduction or induction, or combine strategies.
  Why it matters: This keeps the framework applicable to systems that reason in different ways.

Key terms: knowledge base, reasoning engine, observations, feedback loop, deductive, inductive, reasoning strategies

2.5.1 Limitations of a pure deductive reasoning engine

This section uses a virtual-doctor example to show both the potential and the main limitation of a purely deductive reasoning engine.

• Knowledge requirements for deductive reasoning: Explain how a deductive reasoner could select diagnostic actions from a knowledge base, and why this approach is limited in practice.
  Why it matters: The reasoner’s conclusions depend on the quality and completeness of the knowledge it uses.

Key terms: deductive reasoner, knowledge base, optimal actions

2.5.2 Using inductive reasoning and ML

This section explains how ML-powered inductive reasoning broadens a reasoning engine’s knowledge base and supports predictions from incomplete information.

• Inductive Reasoning Benefits: Explain how ML-powered inductive reasoning addresses limitations of purely deductive reasoning by expanding the knowledge base and allowing inference under uncertainty.
  Why it matters: These are the two ways inductive reasoning extends what a purely deductive engine can do.
• Inductive Reasoner Workflow: Trace the two steps described for an inductive reasoner, from structuring raw data for the knowledge base to using observed patterns for predictions or actions.
  Why it matters: The two-step account connects preparing knowledge with using it for inductive predictions.
• Manual Feature Selection Burden: Describe the role of manually selected knowledge-base features in traditional ML prediction models and why that selection can be difficult.
  Why it matters: Feature selection is a practical difficulty in using traditional ML for this process.

Key terms: inductive reasoning, ML, ontologies, inference under uncertainty, inductive reasoner, structured format, knowledge base, KG, NLP, traditional ML approaches

2.5.3 The role of LLMs in the reasoning engine

This section explains how LLMs help a reasoning engine handle ambiguous inputs and incomplete information, especially when combined with KG-based reasoning.

• Probabilistic reasoning under uncertainty: Explain how an LLM can consider multiple possibilities when information is incomplete, rather than requiring explicit rules or giving one definitive answer. Use the medical example to describe a prioritized diagnostic approach.
  Why it matters: It explains why an LLM can still offer useful suggestions when a rule-based process lacks enough information.
• LLM and KG reasoning roles: Describe the LLM's role when integrated with a KG-based reasoning engine and why the section presents this as useful in real-world conditions.
  Why it matters: It connects the LLM's ability to handle uncertainty to the purpose of combining it with KG-based reasoning.

Key terms: probabilistic reasoning, purely deductive systems, uncertainty, prioritized diagnostic approach, KG-based reasoning engines, reasoning layer, ambiguous inputs, hybrid approach, IASs

2.6 A KG approach to IASs

The section contrasts data-first KG construction with a business-driven approach to building and evolving knowledge graphs for intelligent agent systems.

• Limits of data-first construction: Explain why importing and consolidating available data before identifying user tasks can undermine KG adoption for intelligent agent systems.
  Why it matters: It explains why collecting more data is not, by itself, a sound plan for an intelligent system.
• Business-driven KG construction: Describe how business objectives determine which data to acquire and how to shape the KG for an intelligent agent system.
  Why it matters: Starting with the intended use keeps the graph focused on knowledge the agent actually needs.
• Modeling through deployment phases: Trace what the section says happens after KG acquisition, from modeling and evaluation through product deployment.
  Why it matters: The KG is part of a tested and deployed system, not merely a collected dataset.
• Iterative schemaless KG extension: Explain how later development rounds extend an existing KG without undoing earlier results, and what schemaless flexibility contributes.
  Why it matters: The system can grow to meet new needs while retaining what earlier rounds delivered.

Key terms: bottom-up approach, single source of truth, KG adoption, business understanding, data understanding, purpose-driven approach, domain-specific, knowledge graph, acquisition, modeling phase

2.7 Summary