Part VI of the AIMA book covers learning and various techniques on agents can learn.  The material covers Learning from Observations, Neural Networks, Reinforcement Learning, and Knowledge in Learning.  In the second edition, the chapter in Neural Networks is replaced with Statistical Learning Methods.

I read Chapter 18 of the AIMA book called Learning from Observations.  The Chapter focuses on decision trees and decision lists as some computational learning theory.  The main classification algorithm is the ID3 algorithm for classifying examples and generating decision trees.  The ID3 algorithm is based on Hunt’s Concept Learning System (CLS).  The ID3 algorithm uses information theory for obtaining the decision tree covering most the examples.  I have studied the ID3 algorithm quite extensively since the original application was to classify the winning position in the king and rook versus the king and knight endgame (Quinlan 1983).

Chapter 19 covers Neural Networks (in the second edition this chapter is replaced with Statistical Learning Theory).  The chapter reviews single and multilayer network networks as well as the perceptron, hopfield networks, and boltzmann machines.  Also, the Bayesian Belief Network is discussed.  Although there has been with some successes such as NET talk, hand recognition, and ALVINN.  Still neural networks have not reached a level of success as other implementations of machine learning.

Chapter 20 covers Reinforcement Learning.  This chapter the authors present reinforcement learning by covering the model based approach.  The utility function is maximized based on the state and action yielding the reward.  Unfortunately, the state space in many problems is very learning, and therefore, it is impractical to assign values to a large state space.  Other methods such as temporal difference (TD) methods are used and Q-learning are used to find the optimal policy.  The authors discuss the published successes in games such as Neurogammon, TDgammon, and Othello.  Current game research using reinforcement learning is investigating Go.  Lastly, the authors consider Evolutionary Computing (EC) as a special subset of reinforcement learning since the goal is to find the optimal fitness function.

Finally Chapter 21 covers Knowledge in Learning.  The authors present the case about learning given background knowledge.  The authors discuss the concept of knowledge base inductive learning (KBIL).  First, attribute learning was presented with background knowledge.  The main difference with Chapter 18 is background knowledge is used to narrow the hypothesis space.  Then Explanation Based Learning (EBL) was presented along with some applications.  In addition, the authors discussed an extension to ID3 with background knowledged (RBID3).  Then, Inductive Logic Programming (ILP) was presented.  ILP can solve problems by using Inverse Resolution (IR) or using top down methods like FOIL.  ILP uses horn clauses for the knowledge base and problem definition.  Upon an ILP process based on the examples, a logic program is generated.  ILP takes advantage of relational learning.

Based on my reading experiences, the ID3 algorithm has been successful algorithm since its successors C4.5 and C5 have commercial success.  However, there are other attribute based classification algorithms such as the AQ-15 by Michalski and CN2 by Clark and Niblett.  Also, related to Reinforcement Learning research is Markov Decision Process (MDP) and Partially Observable MDP (POMDP) models, but these concepts are covered in Chapter 17 of Part V.  With ILP, the Lavrac and Dzeroski Inductive Logic Programming: Techniques and Applications book is available online (see the references page for the URL).  There are many active areas of learning research today.