BP网络其实还是前向反馈网络中的一种,她对于处理非线性分类问题有独特的优势。
BP NN can approach to any Non-Linear Mapping Relationship, only if there are enough hide layer & hide nodes.
Structure of BP Network is same as multi-Layer Perceptron Network only their Transfer Functions are different .
但由于使用了复杂的s函数,实现功能时要麻烦的多,但更新权重的基本原理是一致的,同样使用delta study principle.
需要注意的是不同的层数有不同的delta计算方法。得到delta后可以对权重进行更新。
BP的特点:
Advantages of BP NN:
BP NN can approach to any Non-Linear Mapping Relationship, only if there are enough hide layer & hide
nodes.
Learning Algorithm of BP NN is a kind of Global approach method, so it has a good Generalization Capability BP doesn’t depend on mathematical model. I/O relative information is distributed, stored in mass
Link-Weight.
Damage of Individual neuron has a limited affection.
BP has Strong Fault-Tolerance.
Can be applied to Info. Process, Graphic&Pattern.
Recognition, System Control, but not real-time control.
Disadvantages of BP NN:
Since the Cost Function E is a very complex super surface of Link-Weight, so the Convergence is rather
slow.
Super surface usually has multi Extreme Points, the algorithm possible converges to a Local Extreme
Point and can’t obtain the Global Optimization Point.
Principally BP NN can implement very complex mapping relationship, only if there are enough hide layer & hide nodes. But it is hard to determine the number of hide layer & hide nodes.