Summary of all notes: "Principles of Artificial Neural Networks"-Summary of Reading Notes
1. Comparison of neural network and expert system
Knowledge representation
- The original forms of knowledge representation in the expert system are all based on formal symbols, and knowledge is all explicit and descriptive representations;
- The representation of knowledge by neural network is implicit. It uses the topology of the network to represent the relationship between nodes.
Knowledge acquisition
- The main way for the expert system to acquire knowledge is mechanical learning and lecture-based learning, and the acquired knowledge must be deterministic and noise-free, otherwise it will affect the correctness of the system's conclusions;
- Neural network has strong self-organization, self-learning, and self-adaptive capabilities. It can obtain symbolic concepts through learning from digital examples, and can also obtain knowledge from existing knowledge bases or explicit statements of knowledge by domain experts , And can deal with inaccurate, incomplete, and noisy knowledge.
Knowledge reasoning
- The expert system mainly uses logical reasoning. In reasoning, it searches for corresponding knowledge in the knowledge base according to a certain matching algorithm and search strategy. It is usually a deterministic reasoning and a serial processing method. The reasoning efficiency is not high and it is difficult to solve the rules. Conflict resolution problem;
- The reasoning process of the neural network can achieve the effect of uncertain reasoning, and the inherent parallel reasoning ability of the neural network greatly improves the reasoning efficiency and solves the conflict problem in logical reasoning.
Conclusion interpretation
- The expert system can clearly explain the conclusions obtained to the user;
- It is difficult for neural networks to give a clear explanation to the conclusion.
2. Comparison of neural network and analog system
Combination of neural network and simulation system
- Introducing ANN into the fuzzy system
Use ANN as a networked description form of membership functions, fuzzy rules and expansion principles in fuzzy systems. - Introduce the fuzzy principle into the existing ANN.
The neural network structure in the training and working process is regarded as a kind of fuzzy category sign.
The fuzzy principle is applied to the neuron, so that the neuron can behave as a variety of fuzzy operations in function, such as fuzzy intersection, fuzzy weighting, and so on.
Fuzzy preprocessing is performed on the input data of the neural network.
The fuzzy relationship is introduced into the neural network structure and combined with the learning mechanism of the neural network.