"Machine learning" reading notes, Chapter I. INTRODUCTION

"Machine learning" in mechanical engineering press is a very good introductory book, here to organize my study notes

The first chapter, Introduction


1.1 Definition of machine learning:
For a certain type of task T and performance Metric p, if a computer program is self-perfecting with experience E in the performance of P measured on T, then we call this computer program to learn from experience E.


1.2 Designing a learning system:
In general, the training sample is distributed in a similar distribution to the test sample, and learning has the greatest credibility. Most machine learning theories now rely on the assumption that the training sample is consistent with the test sample distribution.
Although we need such assumptions in order to obtain theoretical results, it is also important to remember that this hypothesis is often not tenable in practice.

Checkers Learning Questions:
Task T: Next Checkers
Performance indicator P: Percentage of opponents defeated at world championships
Training Experience E: doing chess with yourself
In order to complete the design of this learning system, we need to choose
1) The exact type of knowledge to learn (objective function)
of the target function
2) Representation of this target knowledge (select feature)
3) A learning mechanism (estimating the training value, adjusting the weight value)
LMS weighted updating Method--minimum Average method
For each training sample <b, Y (b) >
Calculates h (b) using the current weight value
Update for each weight w: w = w + Etah (Y (b)-H (b)) * X (i)


1.2.5 Final Design:
The core components of a learning system ———— four different program modules:
Experiment generator--execution system--appraiser--generalization--Experiment generator
Executing the system, using the learning objective function to solve a given task, we expect its performance to improve as the evaluation function becomes more accurate.
An appraiser, which takes the history as input and outputs a series of training samples of the target function
The generalization, which takes the training sample as input, produces an output hypothesis as its estimation of the objective function
The experiment generator, which takes the current assumptions as input, outputs a new problem for the execution system to explore


1.3 Machine-Learning Perspectives:
Machine learning problems are often attributed to search problems, that is, searching for very large hypothetical spaces to determine best fit to observed data and machine learning assumptions of knowledge
So the task of machine learning is to search for this space and find the hypothesis that best fits the training data.
The LMS algorithm for fitting weights achieves this goal by iterating over the weighted values.


Many chapters of this book provide a search algorithm for hypothetical spaces defined by some basic representations (e.g., linear functions, logical descriptions, decision trees, artificial neural networks)
These different hypothesis representations are suitable for learning different objective functions
For each of these hypothetical representations, the corresponding learning algorithms exert the advantages of different intrinsic structures to organize the search for the hypothetical space.


Throughout, this book runs through this view of learning as a search problem
Thus, the learning method is characterized by search strategy and the intrinsic structure of search space explored by the learner.


Machine Learning Problems:
What algorithms exist to learn the general objective function from the specific training data.
If sufficient training data is provided, the conditions under which the particular algorithm converges to the desired function.
Which algorithm has the best performance for which problems and representations.


How much training data is sufficient.
How to find the confidence degree of the assumed function, the number of training data, and the general relationship between the hypothetical spatial characteristics provided to the learner.


The prior knowledge possessed by the learner is how to guide the process of generalization from the sample.
Are they helpful when prior knowledge is only approximate and correct?


What is the best strategy for choosing an effective follow-up training experience?
How the choice of this strategy will affect the complexity of learning problems.


How to simplify the learning task into one or more function approximation problems.
In another way, the system is trying to learn which functions.
Can the process itself be automated?


How the learner automatically changes the representation to improve the ability to express and learn the objective function.