Reading Notes on Chapters 4–5 of Artificial Intelligence: A Modern Approach

不要问我为什么

    Chapter 4 of "Artificial Intelligence: A Modern Approach" delves into the intricate topic of representing knowledge, a crucial aspect of building intelligent systems. Russell and Norvig introduce various knowledge representation formalisms, including propositional logic (e.g., "P ∧ Q → R" for compound statements), first-order logic (with quantifiers like "∀x" and "∃y"), and semantic networks (graphical representations of entities and relations). The chapter uses the Wumpus World example again to illustrate how logical representations can encode domain knowledge, e.g., "∀x,y [(Cave[x] ∧ AgentAt[y]) → (AgentSmells[y] ↔ Nearby(x,y) ∧ Wumpus(x))]" captures the smell rule. However, the authors highlight the expressiveness vs. tractability trade-off: while first-order logic is powerful, reasoning tasks like satisfiability become NP-complete. They discuss resolution theorem proving as a method for logical entailment, but point out its pitfalls, such as the exponential growth of clauses in complex domains. Frame representations and scripts (e.g., restaurant dining script) are introduced as ways to manage default knowledge and commonsense reasoning, yet they struggle with exceptions and novel situations. A deep insight here is the role of ontology in knowledge representation: how should concepts and relationships be structured to facilitate reasoning? For instance, the CYC project aims to encode vast amounts of common sense knowledge, but faces challenges of consistency and scale. This raises questions about the limits of formal logic for capturing human-like knowledge, especially in ambiguous or context-dependent scenarios.  
    Chapter 5 shifts focus to planning, a critical component of goal-driven behavior in AI agents. The chapter distinguishes between classical planning (e.g., the blocks world with deterministic actions and a fixed goal state) and situational calculus (handling dynamic environments with action preconditions and effects). STRIPS (Stanford Research Institute Problem Solver) is presented as a key planner that uses a state space representation and operators to achieve goals. For example, in a logistics problem (loading trucks), STRIPS operators like "Load(x,y)" and "Unload(x,y)" can be used to plan efficient routes. However, the chapter discusses the challenges of planning under uncertainty, such as in the medical diagnosis domain where symptoms can be ambiguous. Partial order planning and contingent planning (e.g., with "IF-THEN" conditions) are introduced as extensions, but face issues of computational complexity and incomplete information. The utility theory approach to planning, which assigns values to outcomes and selects actions based on expected utility, raises ethical questions: how should utilities be defined in contexts with moral dilemmas or conflicting stakeholder interests? The chapter also explores hierarchical task networks (HTNs) and reinforcement learning (RL) for planning, highlighting the trade-offs between top-down planning and bottom-up learning. A key takeaway is the tension between optimality (finding the best plan) and efficiency (computing plans quickly in real-time), especially in dynamic environments like autonomous driving where fast decisions are crucial.  

Depth of Reflection:  
    Chapters 4 and 5 collectively highlight the challenges of knowledge representation and planning in AI, which are deeply intertwined with cognitive science, philosophy, and ethics. The expressiveness vs. tractability dilemma in knowledge representation reflects broader debates about the limits of formal logic for capturing human-like reasoning, particularly in areas like commonsense knowledge and context-dependent understanding. The CYC project’s ambitions to encode vast knowledge highlight the difficulties of scaling formal representations to real-world complexity. Similarly, the challenges of planning under uncertainty (e.g., medical diagnosis) reflect the tension between idealized models and messy reality. The utility theory approach to planning raises profound ethical questions: how should AI systems balance different stakeholder values, especially in high-stakes domains like healthcare or finance? Moreover, the contrast between classical planning and reinforcement learning (RL) underscores the broader debate between symbolic AI (rule-based systems) and connectionist AI (learning from data). While classical planners excel at logical reasoning and goal-driven behavior, RL systems show promise in learning from experience in complex environments. However, RL faces challenges of exploration-exploitation trade-offs, long-term planning horizons, and interpretability of learned policies. Ultimately, these chapters emphasize the need for hybrid approaches that combine logical reasoning, probabilistic models, and learning algorithms to build truly intelligent systems that can navigate the complexities of the real world.