AUTOMATED DESIGN OF AN ARTIFICIAL NEURON FOR FIELD-PROGRAMMABLE GATE ARRAYS BASED ON AN ALGEBRA-ALGORITHMIC APPROACH

Volume 67, Issue 5, 2022, pages 61-72

DOI: http://doi.org/10.34229/2786-6505-2022-5-6

Download

Anatoliy Doroshenko, Institute of Software Systems of National Academy of Sciences of Ukraine, Kyiv, doroshenkoanatoliy2@gmail.com

Volodymyr Shymkovych, National Technical University of Ukraine «Igor Sikorsky Kyiv Polytechnic Institute», shymkovych.volodymyr@gmail.com

Tural Mamedov, Institute of Software Systems of National Academy of Sciences of Ukraine, Kyiv, tural.mamedov1@gmail.com

Olena Yatsenko, Institute of Software Systems of National Academy of Sciences of Ukraine, Kyiv, oayat@ukr.net


ABSTRACT

Neural network control systems are a high-tech branch of control theory and belong to the class of nonlinear dynamic systems. High speed due to the parallelization of input information combined with the ability to train neural networksmakes this technology very attractive for creating control devices in automaticsystems. The high-speed operation of networks in real time is provided by implementing them on field-programmable gate arrays. An example of the hardware implementation of neural networks is the design of an artificial neuron andits nonlinear activation functions on an FPGA. The technology of developingapplications for FPGAs is based on the presentation of the algorithm in thehardware description language, for example, VHDL, and the automatic translation of this description into a specification at the level of logic tables and otherfunctional components of the FPGA. Programming in the VHDL language isquite complex, so the question arises about the development of special softwareautomation tools that would allow the efficient generation of high-performancecode. The paper proposes the facilities of automated design and generation ofprograms for FPGAs based on the algebra of algorithms. The developed toolsare applied for the automated design of an artificial neuron. A method of constructing an artificial neuron with a sigmoidal activation function on an FPGA isdeveloped, which differs from similar approaches in that the coefficients of thepiecewise linear approximation of the activation function are stored in memoryonly for positive or only for negative values of the arguments. This made it possible to optimize the number of utilized computing resources and increase theperformance of the neural network. The developed approach is applied to thedevelopment of a system with a neural controller for balancing a ball on a platform implemented on an FPGA.

Keywords: automated design, algebra of algorithms, approximate computing, control system, field-programmable gate array, neural network, program generation.


REFERENCES

  1. Barron A.R. Universal approximation bounds for superposition of a sigmoidal function. IEEETransactions on Information Theory. 1993. Vol. 39, N 3. P. 930–945. http://doi.org/https://doi.org/10.1109/18.256500.
  2. Jhang J.-Y., Tang K.-H., Huang C.-K., Lin C.-J., Young K.-Y. FPGA implementation of a functional neuro-fuzzy network for nonlinear system control. Electronics. 2018. Vol. 7, N 145. P. 1–22.http://doi.org/10.3390/electronics7080145.
  3. Costarelli D., Spigler R. Approximation results for neural network operators activated by sigmoidal functions. Neural Networks. 2013. Vol. 44. P. 101–106. http://doi.org/10.1016/j.neunet.2013.03.015.
  4. Godse A.P., Godse D.A. VHDL programming: concepts, modeling styles and programming. Seattle: Amazon Digital Services LLC, 2020. 206 p.
  5. Doroshenko A., Yatsenko O. Formal and adaptive methods for automation of parallel programsconstruction: emerging research and opportunities. Hershey : IGI Global, 2021. 279 p. http://doi.org/10.4018/978-1-5225-9384-3.
  6. Pohl C., Paiz C., Porrmann M. vMAGIC — automatic code generation for VHDL. InternationalJournal of Reconfigurable Computing. 2009. Vol. 2009, Article ID 205149. P. 1–9. http://doi.org/10.1155/2009/205149.
  7. Martín P., Bueno E., Rodríguez Fco. J., Machado O., Vuksanovic B. An FPGA-based approach tothe automatic generation of VHDL code for industrial control systems applications: a case studyof MSOGIs implementation. Mathematics and Computers in Simulation. 2013. Vol. 91. P. 178–192.http://doi.org/10.1016/j.matcom.2012.07.004.
  8. de Bulnes D.R.F., Maldonado Y. VHDL code generation as state machine from a data flow graph.Proc. 2016 IEEE International Autumn Meeting on Power, Electronics and Computing(ROPEC 2016) (09–11 November 2016, Ixtapa, Mexico). New York : IEEE, 2016. P. 1–6. http://doi.org/10.1109/ROPEC.2016.7830518.
  9. Benáček P., Puš V., Kubátová H., Čejka T. P4-To-VHDL: automatic generation of high-speed input and output network blocks. Microprocessors and Microsystems. 2018. Vol. 56. P. 22–33.http://doi.org/10.1016/j.micpro.2017.10.012.
  10. Goel K., Arun U., Sinha A. K. An efficient hardwired realization of embedded neural controlleron System-On-Programmable-Chip (SOPC). International Journal of Engineering Research &Technology. 2014. Vol. 3, N 1. P. 276–284.
  11. Singh S., Sanjeevi S., Suma V., Talashi A. FPGA implementation of a trained neural network.IOSR Journal of Electronics and Communication Engineering. 2015. Vol. 10, N 3. P. 45–54.
  12. Mohammadi A., Ryu J.-C. Neural network-based PID compensation for nonlinear systems: ballon-plate example. International Journal of Dynamics and Control. 2020. Vol. 8. P. 178–188.http://doi.org/10.1007/s40435-018-0480-5.
  13. Shaheer M., Hashmi H., Khan S., Atif M., Shabbir Z., Ali A., Kamal K., Zafar T., Awan A. Control of a ball-bot using a PSO trained neural network. Proc. 2nd International Conference onControl, Automation and Robotics (ICCAR 2016) (28–30 April 2016, Hong Kong, China). NewYork : IEEE, 2016. P. 24–28. http://doi.org/10.1109/ICCAR.2016.7486692.
  14. Han K., Tian Y., Kong Y., Li J.,. Zhang Y. Tracking control of ball and plate system using an improved PSO on-line training PID neural network. Proc. 2012 IEEE International Conference onMechatronics and Automation (5–8 August 2012, Chengdu, China). New York : IEEE, 2012.P. 2297–2302.
  15. Doroshenko A., Shymkovych V., Yatsenko O., Mamedov T. Automated software design forFPGAs on an example of developing a genetic algorithm. Proc. 17th International Conference«ICT in Education, Research and Industrial Applications. Integration, Harmonization andKnowledge Transfer» (ICTERI 2021) (28 September – 2 October 2021, Kherson, Ukraine). 2021.P. 74–85.
  16. Camponogara E., Nazari L.F. Models and algorithms for optimal piecewise-linear function approximation. Mathematical Problems in Engineering. 2015. Vol. 2015, Article ID 876862.P. 1–9. http://doi.org/10.1155/2015/876862.
  17. Shymkovych V., Samotyy V., Telenyk S., Kravets P., Posvistak T. A real time control system forbalancing a ball on a platform with FPGA parallel implementation. Technical Transactions. 2018.Vol. 5. P. 109–117. http://doi.org/10.4467/2353737XCT.18.077.8559.