"AI Tutorial (Zhang Hisen)" (iv)

Six, machine learning

1. Overview

Machine learning is the automatic acquisition of knowledge by computers.

The first thing to talk about is machine learning . There are three research goals for machine learning:

1) cognitive model of human learning process.

2) Universal Learning algorithm.

3) construct a task-oriented special learning system (engineering objectives).

Then we talk about the significance of studying machine learning . Machine learning speed, easy to accumulate knowledge, learning results easy to spread.

Then we talk about the history of machine learning . Since the 1950s, it can be divided into four stages, the research contents are as follows:

1) The study of neuron model and decision theory.

2) Symbolic concept acquisition research.

3) Knowledge enhancement and domain-specific learning system.

4) Connect Learning research.

Finally, the main strategies and research status of machine learning are discussed. The reasoning methods used in learning are called learning strategies. Several basic strategies are:

1) Mechanical Learning. Also known as memory learning, is the simplest learning strategy.

2) Impart learning. Also known as guided learning or pointing learning.

3) Deductive learning.

4) Inductive learning. According to its teacher guidance, can be divided into example learning (also known as concept acquisition, a set of positive and inverse examples) and observation and discovery Learning (also known as the generalization of description).

5) Analogy learning.

2. Basic model of machine learning system

in this basic model, 4 basic components are included.

1) environment. Refers to the source of external information of the system, providing the learning system with the material or information necessary to acquire knowledge of the relevant objects.

2) Learning links. Get knowledge of related objects from the environment-provided and required variances, and store that knowledge in the knowledge base.

3) Knowledge Base. Store the knowledge learned from the learning process. In choosing the method of knowledge representation, the criteria to be considered are: the strength of expression ability, the size of reasoning difficulty, the difficulty of modification, and the ease of expansion.

4) Implementation of the process. Is the core of the entire learning system. The knowledge-solving problem learned in the knowledge base is applied, and the evaluation results are fed back to the learning link, which facilitates the further study of the system. The factors that affect the learning link are: The complexity of the problem, the feedback information, the transparency of the execution process.

3. Mechanical Learning

First, we talk about the process of mechanical learning . That is, the system remembers the problem and its solution for every problem that the executing agency solves.

Then we discuss the problem that the mechanical learning system should consider .

1) storage structure.

2) The stability of the environment and the applicability of stored information.

3) The trade-offs stored in the calculation. Two methods: Cost-benefit analysis method and recently unused substitution method (selective waiver method).

4. Teach-In learning

is a comparative learning method, which can be used for knowledge acquisition of expert system. The following 5 steps are included:

1) requirements. Ask the expert to make a recommendation.

2) explanation. The transfer of expert advice into an internal representation is a matter of knowledge representation.

3) Practical. The process of information transformation, the transfer of abstract ideas into specific knowledge. is a core step.

4) Join the Knowledge base. Add new knowledge to the knowledge base.

5) evaluation.

5. Analogy Learning

First, we talk about learning new concepts . There are three issues to be aware of:

1) Learners must have certain knowledge, otherwise they will not achieve the effect of learning.

2) There must be similar attributes between the errors used in the analogy.

3) Direct comparisons between attributes and their values often do not explain the problem, only after the abstract attributes can reflect the nature of the analogy.

Then we discuss the solving method of the learning problem . There are two kinds: the change analogy method and the derivation analogy method. There are two processes in conversion analogy learning: the process of recalling and the process of transformation.

6. Inductive learning

According to its teacher guidance can be divided into examples of learning and observation and discovery learning two forms.

First, we talk about example learning . Also known as sample learning or through sample learning. The examples given in advance are divided into positive and inverse cases by teachers. The general knowledge given should be able to explain all the given positive examples and exclude all given counter-examples. In general, positive and inverse examples are provided by the information source. Information sources are divided into: teachers who already know the concept, learners themselves, outside the learner's external environment. Instance learning is one of the most well researched and fruitful branches in machine learning field. Two spatial models in the case study:

1) Example space: A collection of training examples provided to the system. The questions to consider are: the quality of the teaching example, the organization of the example space, and the search method.

2) Rule space: a law of things. Two issues related to this are the requirements for the rules space and the search method for the rule space.

Several inductive reasoning methods are commonly used:

1) The constant quantization as a variable. By the concrete to the general evolution or induction.

2) Remove the condition. Remove extraneous sub-conditions (partial constraints).

3) Increase the selection. Increase the disjunction and expand the scope. Commonly used methods are: The first part of the analysis method and internal analysis.

4) Curve fitting.

The methods for searching the rule space are:

1) Deformation space method. A data-driven approach.

2) Improve the Assumption method. A data-driven approach.

3) generated in the test method. A model-driven approach.

4) Scenario Example method. A model-driven approach.

The advantage of the data-driven approach is that it is possible to gradually accept teaching examples to learn. The advantage of model-driven method is that it has good anti-jamming characteristics.

Then we talk about observation and discovery learning . Inductive learning without Teacher's guidance is divided into:

1) Observe learning. Used for conceptual clustering of examples to form conceptual descriptions.

2) Discover learning. Can be divided into empirical discoveries (laws and laws found in empirical data) and knowledge discovery (discovering new knowledge from observed examples).

7. Interpretation-Based Learning

belongs to a kind of deductive learning method. The general descriptive framework is:

1) Given: domain knowledge dt; target concept TC; training example te; operability criterion oc.

2) Find out: Sufficient conditions for the TC to meet OC.

First of all, we talk about the working principle based on explanation learning .

1) construct the explanatory structure. To prove why the instance provided to the system is an instance of satisfying the concept of the target.

2) General processing of the resulting explanatory structure to obtain general knowledge.

Then, for example .

Finally, the perfection of domain knowledge . As far as possible to provide a sound field knowledge.

8. ID3 Decision Tree algorithm

the ID3 (interative dichotomic Version3) algorithm is an example learning approach, formerly known as the Conceptual Learning System (CLS).

First, we discuss the ID3 algorithm . The process of main algorithm and the process of achievement are illustrated.

Then we discuss the example calculation .

Finally, the characteristics of the ID3 algorithm . Ideal for non-incremental learning tasks, not for incremental learning tasks.

Copyright NOTICE: This article for Bo Master original article, without Bo Master permission not reproduced.

"AI Tutorial (Zhang Hisen)" (iv)