Implementation of neural networks in the method of multilevel component circuits
The paper analyzes the features of representing artificial neural networks in Simulink and SimInTech. Examples of visual schemes (models) built in these modeling environments using neural network blocks are given. The following shortcomings of such representations are the lack of mechanisms: for ca...
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Saint Petersburg National Research University of Information Technologies, Mechanics and Optics (ITMO University)
2023-12-01
|
| Series: | Naučno-tehničeskij Vestnik Informacionnyh Tehnologij, Mehaniki i Optiki |
| Subjects: | |
| Online Access: | https://ntv.ifmo.ru/file/article/22498.pdf |
| Summary: | The paper analyzes the features of representing artificial neural networks in Simulink and SimInTech. Examples of
visual schemes (models) built in these modeling environments using neural network blocks are given. The following shortcomings of such representations are the lack of mechanisms: for carrying out structural optimization of neural
networks, for combining them into ensembles, for training them synchronously with the simulation of the object model.
It was noted that there are difficulties in using other tools, such as specialized Python libraries (Keras, PyTorch, etc.),
the NeuroGenetic Optimizer (BioCompSystems) for building neural network control models. A method is shown to
implement the representation of neural networks in the formalism of the method of multilevel component circuits,
according to which the construction of models of an object and a control system is carried out in a visual language from
ready-made blocks (components) with directional and non-directional connections. A technique has been developed
for multilevel representation of neural network control models, which allows them to be combined with other tools
of the component circuit method. Two options for representing neural networks are proposed: with an encapsulated
structure and with a component structure. The first version of the representation is characterized by the compactness of
the representation of the control model, the possibility of automated variation and optimization of the structure of the
neural network, and the possibility of changing the structure of the network during the executing of the model within a
computational experiment (scenario). The second option has the ability to perform detailed debugging and research of
the network learning process, and the ability to construct a network of any structural complexity. The paper describes the
main developed components with their connections: a neural network, a training block, an ensemble unit (bagging), a
block for reading data from a file, a sampling block, a neural network layer (input, hidden, output). A multilevel computer
model of the uncontrolled flight of a body (target) and the controlled flight of a projectile is presented as an example
to illustrate the operation of the developed components to solve the problem of controlling a projectile to hit the target.
The developed component libraries can be used as part of the MARS modeling environment to build multilevel control
systems for objects of a multiphysics nature. |
|---|---|
| ISSN: | 2226-1494 2500-0373 |