Russell Norvig Chapter 01: Introduction
Metadata
Title: Introduction
Number: 1
Exercises: https://aimacode.github.io/aima-exercises/intro-exercises/
Coverage
1.1 What is AI? - Approaches to definition
1.2 Foundations of AI - disciplinary foundations
1.3 History of AI - phases of AI research and applications
1.4 State of the Art
1.5 Risks and Benefits
Core Ideas
Key Quotes
The field of artificial intelligence, or AI, is concerned with not just understanding but also building intelligent entities - machines that can compute how to act effectively and safely in a wide variety of novel situations.
Introduction, p. 19
What is AI?
Historically researchers have pursued different versions of AI, some have defined intelligence in terms of fidelity to human performance. Others prefer an abstract, formal definition based on rationality. Some consider intelligence a property of internal thought processes and reasoning, others focus on intelligent behaviour.
This gives us two dimensions human vs. rational, thought vs. behaviour. There are four combinations here, and research programs exist in all four. RN then present the main approaches according to these combinations:
1. Acting Humanly (The Turing Test)
Epitomised by the Turing Test. To pass the test the machine needs: NLP, Knowledge Representation, Automated Reasoning, and Machine Learning.
Turing was not interested in the physical simulation of a person, later researchers have propsed a total Turing test which requires interaction with objects and people in the world. Here a robot would need in addition computer vision and robotics.
These six faculties compose most of AI, but the Turing test itself does not occupy most AI research.
2. Thinking Humanly (Cognitive modeling)
Cognitive science brings together computer models from AI and experimental techniques from psychology to construct precise and testable theories of the human mind. It is a huge field in itself, with a lot of cross-fertilization with AI research.
In the early days of AI there was a lot of confusion between the approaches - with authors claiming that if an algorithm performed well on a task it is therefore a good model of human performance, or vice versa. Now more careful to separate the two.
3. Thinking Rationally (Laws of thought)
Draws on formal logic to codify “right thinking”.
By 1965, programs could, in principle, solve any solvable problem described in logical notation. The so-called logicist tradition within artificial intelligence hopes to build on such programs to create intelligent systems.
Introduction, p. 21
But logic conventionally requires knowledge of the world that is certain. In situations of uncertainty, probability theory is needed to fill the gap. Thinking rationally also does not produce intelligent behaviour. We still miss a theory of rational action.
4. Acting Rationally (Rational agent)
An agent is something that acts (from the latin agere, to do). All computer programs do something, but computer agents are expected to do more - operate independently, perceive their environment, persist over time, create and pursue goals, adapt to change.
A rational agent is one that acts so as to achieve the best outcome, or the best expected outcome in the case of uncertainty.
The rational-agent approach to AI has two advantages over the other approaches. First, it is more general than the “laws of thought” approach because correct inference is just one of several possible mechanisms for achieving rationality. Second, it is more amenable to scientific development. The standard of rationality is mathematically well defined and completely general. We can often work back from this specification to derive agent designs that provably achieve it - something that is largely impossible if the goal is to imitate human behaviour or thought processes.
Introduction, p. 22
As such this approach has prevailed for most of the field’s history:
In a nutshell, AI has focused on the study and construction of agents that do the right thing. What counts as the right thing is defined by the objective that we provide to the agent. This general paradigm is so pervasive that we might call it the standard model. It prevails not only in AI, but also in control theory, where a controller minimizes a cost function; in operations research, where a policy maximizes a sum of reward; in statistics, where a decision rule minimizes a loss function; and in economics, where a decision maker maximizes utility or some measure of social welfare.
Introduction, p. 22
However it is not alway feasible in complex environments to compute the exact optimal outcome, the computational demands are too high. So even if that is the theoretical goal we need to deal with the issue of limited rationality, acting appropriately when we cannot do all the computations we wish.
5. Beneficial Machines
RN think that that the standard model is not the right one for the long run, because it assumes we supply a fully specified objective to the machine. This makes sense for an artificially defined task like playing chess, but in the real world it becomes harder to specify the objective completely and correctly. EG self-driving cars, what exactly is their objective?
The problem of achieving agreement between our true preferences and the objective we put into a machine is called the value alignment problem. It has become a critical issues as AI applications enter the real world.
Take chess for example, if the machine treated winning as its sole objective, and could operate outside the confines of a board, it might resort to blackmail or coercion to win. This is a logical consequence of the goal’s definition. We want machines to pursue our objectives, not their objectives, and agents that are provably beneficial to humans.
Foundations of AI
The chapter reviews briefly the main disciplines that have contributed to AI research (summaries from p. 52):
Philosophy, p. 24 - made AI conceivable by suggesting that the machine is in some ways like a machine, that it operates on knowledge encoded in some internal language, and that thought can be used to choose among actions.
Mathematics, p. 26 - provided the tools to manipulate statements of logical certainty as well as uncertain, probabilistic statements. Set the groundwork for understanding computation and reasoning about algorithms.
Economics, p. 27 - formalized the problem of making decisions that maximize the expected utility of the decision maker.
Neuroscience, p. 29 - discovered some facts about how the brain works and the ways in which it is similar to and different from computers.
Pyschology, p. 30 - adopted the idea that humans and animals can be considered information processing machines.
Computer engineering, p. 32 - provided the more powerful machines that make AI applications possible, and software engineers to make them usable.
Control theory and cybernetics, p. 33 - deals with designing devices that act optimally based on feedback from the environment. Initially using very different mathematical tools from AI, but coming closer together now.
Linguistics, p. 34 - showed that language use fits the model of the psychologists.
History of AI
The chapter reviews briefly the historical developments in the field, dividing it into a number of phases:
Inception (1943 - 56), p. 35
Early enthusiasm, great expectations (1952 - 69), p. 36
Dose of reality (1966 - 73), p. 39
Expert Systems (1969 - 86), p. 40
Return of neural networks (1986 - present), p. 42
Probabilistic reasoning and ML (1987 - present), p. 42
Big Data (2001 - present), p. 44
Deep Learning (2011 - present), p. 44
State of the Art and Challenges
The chapter concludes by reviewing the state of the art across a range of tasks, pointing to the https://aiindex.stanford.edu/ as a good source for updates.
It then looks at the ‘risks and benefits’ of AI, looking at how it could all go horribly wrong.