Topic: Learning notes

ArtInt Reading Notes: Intelligent Agents and the AI Basics I Mixed Up

Notes on ArtInt 3e sections 1.1–1.2: agents, computability, and corrections to my understanding of decision trees, reinforcement learning, early language systems, and text generation.

After reading sections 1.1–1.2 of Artificial Intelligence: Foundations of Computational Agents, third edition, I found several distinctions I needed to revisit: artificial versus fake, equivalent models versus a thesis, invention versus adoption, and text generation versus verification.

These notes preserve my current understanding, the corrections, and what I still need to practice. The main scope is the opening two sections. Material on MDPs, conditional independence, and tokens is marked as an extension. The scenarios are my own learning examples, not textbook exercises or reports of systems I have implemented.

Start with the agent’s actions

In section 1.1, an agent acts in an environment. A computational agent’s decisions can be explained through implementable basic operations. Intelligent behavior involves goals, limits, consequences, learning, and adaptation; copying the brain is not a requirement of this definition.

I use a work scheduler to make this concrete:

This example helps me separate a prediction from a decision. Estimating that a task takes two hours provides information. Choosing which task to do first also requires goals and constraints.

I also distinguish research from application: executable models can test hypotheses about intelligence, while engineering aims to build useful agents. Collaboration between people and systems belongs in this picture too. Section 1.1

Nine distinctions I needed to correct

This is my review list for catching the same mistaken inferences later.

The conceptual distinctions come from 1.1; the technical and historical distinctions from 1.2.

Artificial, fake, culture, and learning

Saying something is human-made identifies its origin. It does not settle whether the thing really has a particular property. My review habit is to ask what an adjective classifies before using it to reach another conclusion.

Culture and learning should not be separated either. The text describes biology, culture, and lifelong learning as interacting sources of human intelligence. Section 1.1.2

My example is learning the concept of a deadline. Someone teaches me the word and its use; I then apply it to organizing work. Receiving a cultural concept and learning to use it can be parts of the same process.

Church–Turing: what does the equivalence proof establish?

This is where I most needed to slow down.

Turing machines and lambda calculus have formal definitions, and their equivalence in computable functions can be proved. The Church–Turing thesis connects the informal idea of an effective procedure with formal computability. Equivalence between two models does not by itself establish that they cover every effective procedure, much less prove that all thought is algorithmic. Further explanation

I now ask three separate questions:

1. Model capability: Can this formal model compute the function?
2. Scope of the thesis: Does every effective procedure fall within that capability?
3. Practical cost: Can I afford the time and memory needed?

The third question matters for real work. A computable method might still miss the deadline. “Possible in principle” does not settle whether I can use it.

Connecting machine learning, planning, and reinforcement learning

My organizing aid is a workflow: obtain information, choose actions, and use outcomes to improve. This is a learning device, not an architecture every AI system must follow.

Decision trees can learn from data; neural networks are one family of machine learning tools. The text separates prediction from choosing actions and introduces reinforcement learning as decision-theoretic planning with learning. Section 1.2

For my scheduler example, estimating durations differs from choosing an order under deadlines. To discuss learning, I also need to explain what gets updated from which outcomes. The mere fact that a system adjusts something does not establish that it uses reinforcement learning.

Extension: MDPs represent uncertain sequential decisions

An MDP describes states, actions, transition probabilities, and rewards. Under the Markov assumption, the current state and action suffice for the next-state distribution without additional dependence on earlier history. Section 12.5

My example is a robot that can slip. Choosing to move forward does not guarantee success. Planning must account for possible outcomes. If battery level affects those outcomes but the state records only position, the representation may be insufficient.

For now, I am learning what the model is for. I have not implemented an MDP solver, and recognizing its components is not yet evidence that I can model a new problem.

Extension: conditional independence needs a condition

Belief networks use conditional independence to describe probabilistic structure. Section 9.3

In my simplified example, R represents rain, and A and B indicate wet ground at two locations. If A and B are independent given R, the joint distribution factors as:

P(R, A, B) = P(R) × P(A | R) × P(B | R)

The assumption matters. If both locations also share a sprinkler, knowing whether it rained may be insufficient. I need to examine the assumption before using this factorization.

Remember historical systems by their tasks

These associations follow section 1.2. My next review will hide the names and recall them from the tasks, then reverse the exercise. Familiar spelling is not enough.

For Logic Theorist, I start with the tiny inference from P and “if P then Q” to Q. For SHRDLU, I imagine an instruction to place one block on another. Both are my teaching simplifications, not full descriptions of the original systems.

Check generation and truth separately

I confused a generation mechanism with a truth proof. Sequence models can generate subsequent tokens from context; a token may be a character, a word, or part of a word. Generating text does not require proving each sentence true. Extension: section 8.5

My practical correction is to extract checkable claims from an explanation. If a paragraph says a system “first invented” a technique, I can check the date, original work, and actual claim separately. Fluent prose cannot perform those checks for me.

What I want to explain on my next review

• How does a scheduler turn predictions into actions, and what limits it?
• Why does model equivalence fail to prove that all thought is algorithmic?
• Can I identify states, actions, and uncertain outcomes in a new simple scenario?
• Can I state the condition in a conditional-independence claim and identify factors that could break it?
• Can I identify which claims in generated text need verification?

These notes preserve my corrected understanding. I still need to explain the ideas through new examples, without relying on the original choices.

Reading sources

• ArtInt 1.1: agents, intelligent behavior, and natural intelligence.
• ArtInt 1.2: computation, technical approaches, and early systems.
• Church–Turing thesis: effective procedures and formal models.
• Extensions: MDPs, belief networks, and sequence models. These links do not imply that I have completed the chapters or implemented their methods.