The previous chapter discussed how an agent perceives and acts, but not how its goals affect its actions. An agent could be programmed to act in the world to achieve a fixed goal or set of goals, but then it would not adapt to changing goals, and so would not be intelligent. An intelligent agent needs to reason about its abilities and goals to determine what to do.

An agent that is not omniscient cannot just plan a fixed sequence of steps, as was assumed in Chapter 6. Planning must take into account the fact that an agent in the real world does not know what will actually happen when it acts, nor what it will observe in the future. An agent should plan to react to its environment.

This chapter shows how an intelligent agent can perceive, reason, and act over time in an environment. In particular, it considers the internal structure of an agent. As Simon points out in the quote above, hierarchical decomposition is an important part of the design of complex systems such as intelligent agents.

The previous chapter assumed that the input were features; you might wonder where the features come from. The inputs to real-world agents are diverse, including pixels from cameras, sound waves from microphones, or character sequences from web requests. Using these directly as inputs to the methods from the previous chapter often does not work well; useful features need to be created from the raw inputs.

In the example from Pearl (above), mud and rain are correlated, but the relationship between mud and rain is not symmetric. Creating mud (e.g., by pouring water on dirt) does not make rain. However, if you were to cause rain (e.g., by seeding clouds), mud will result. There is a causal relationship between mud and rain: rain causes mud, and mud does not cause rain.

What should an agent do when there are other agents, with their own goals and preferences, who are also reasoning about what to do? An intelligent agent should not ignore other agents or treat them as noise in the environment. This chapter considers the problems of determining what an agent should do in an environment that includes other agents who have their own utilities.

This chapter is about how to represent individuals (things, entities, objects) and relationships among them. As Baum suggests in the quote above, the real world contains objects and compact representations of those objects and relationships can make reasoning about them tractable. Such representations can be much more compact than representations in terms of features alone.

In the machine learning and probabilistic models presented in earlier chapters, the world is made up of features and random variables. As Pinker points out, we generally reason about things. Things are not features or random variables; it doesn’t make sense to talk about the probability of an individual animal, but you could reason about the probability that it is sick, based on its symptoms.

Instead of reasoning explicitly in terms of states, it is typically better to describe states in terms of features and to reason in terms of these features, where a feature is a function on states. Features are described using variables. Often features are not independent and there are hard constraints that specify legal combinations of assignments of values to variables.

In Chapters 7 and 8, learning was divorced from reasoning. An alternative is to explicitly use probabilistic reasoning, as in Chapter 9, with data providing evidence that can be conditioned on. This provides a theoretical basis for much of machine learning, including regularization and measures of simplicity.

This chapter starts with the state-of-the-art in deploying AI for applications, then looks at the big picture in terms of the agent design space (page 21), and speculates on the future of AI. By placing many of the representations covered in the agent design space, the relationships among them become more apparent.

This book is about artificial intelligence (AI), a field built on centuries of thought, which has been a recognized discipline for over 60 years. As well as solving practical tasks, AI provides tools to test hypotheses about the nature of thought itself. Deep scientific and engineering problems have already been solved and many more are waiting to be solved.