COMPUTATIONAL CYBERNETICS
![]() |
Andrey V. Makarenko, Head of Laboratory No. 77 |
The Laboratory is one of the youngest in the Institute. It was established in April 2017 when reorganizing the Institute’s scientific structure to respond adequately to recent challenges of control science. Cand. Sci. (Eng.) Andrey V. Makarenko was appointed its Head.
Computational cybernetics, the main research area of Laboratory No. 77, integrates methods and approaches of two scientific disciplines: cybernetics (the science of control in the broad sense) and computational intelligence (a branch of artificial intelligence with machine learning as the central paradigm). Here, the matter concerns the weak form of artificial intelligence (superadaptive systems) rather than the strong one (artificial general intelligence).
The Laboratory conducts fundamental research at the intersection of two scientific fields:
- nonlinear dynamics (a basis for estimating complex dynamic processes and identifying complex dynamic systems),
- machine learning (a basis for analyzing comprehensively large arrays of weakly structured empirical data and generating descriptive, explanatory, and predictive models over such arrays).
As demonstrated by Laboratory’s research, these scientific fields have a significant constructive interpenetration. For example, the solution of separate (often applied) nonlinear dynamics problems by deep neural networks turns out much more effective compared to classical methods and approaches. On the other hand, using nonlinear dynamics methods in the design, training, and analysis of deep neural networks significantly improves the quality and stability of their operation.
Laboratory’s R & D projects are performed in two key specializations for the Institute:
- complex analysis of data with a complex irregular structure (the observability of controlled objects is estimated);
- mathematical modeling of complex time-varying nonlinear systems (processes) with dominating chaotic behavior (controlled objects are identified, and (or) their dynamics are predicted).
The Laboratory cultivates the rejection of “snobbery” (the existence of an “ideal and the only correct method of solving control problems). The solution always comes from the problem statement with the integration of approaches and methods based on system engineering principles. Subject-oriented experts are necessarily included in the temporary working (project) groups. As a rule, such an organization of work creates the prerequisites for significant breakthroughs in the scientific topics under consideration and a strong competitive advantage in solving complex applied problems: control of heterogeneous and hierarchical systems (for each level, scale, or area) requires their effective mathematical description and a good understanding of the problem details.
Mathematical statistics and probability theory (including the theory of random processes) is another scientific discipline figuring in almost all endeavors of the Laboratory. This can be explained as follows. On the one hand, Laboratory’s employees deal with real objects, and everything in the real world contains errors, noises, skips, outliers, etc. On the other hand, many methods for studying complex systems and (or) processes are probabilistic (e.g., Monte Carlo simulations, training of deep neural networks, and some analysis methods for chaotic discrete mappings). Statistical estimates are needed to judge such seemingly simple things as “there is an effect” or “there is no effect” (“there is a difference” or “there is no difference”). Correct and well-posed conclusions cannot be obtained without testing statistical hypotheses and corresponding probabilistic models.
Note the following results of Laboratory’s employees that are of theoretical importance for the analysis of discrete mappings:
- The T-synchronization of chaotic oscillations was introduced. It generalizes several well-known types of synchronization of chaos. An essential aspect is that T-synchronization allows examining the temporal structure of the synchronism of chaotic systems in a closed form.
- A new type of bifurcations, the so-called TQ-bifurcations, was discovered in discrete mappings. They are associated with a qualitative change in the trajectories of dynamic systems in the extended state space.
- A new approach to estimating the complexity of discrete real sequences, the so-called TQ-complexity, was introduced. An essential aspect is that this measure of complexity is algorithmically realizable and has low computational complexity.
According to the existing stereotypes, deep neural networks effectively operate only on data with pronounced structural patterns. In contrast, Laboratory’s employees study the application of neural networks to processing chaotic and random processes and fields.
Two interesting results were obtained in this area:
- As established, convolutional deep neural networks are constructively applicable to chaos identification in signal processing problems. As shown, deep convolutional neural networks can directly estimate the Lyapunov largest exponent for chaotic systems based on the observed trajectory realization.
- The decision mechanism of deep convolutional neural networks was investigated in the random sequence classification problem. A good coincidence between the analytical and numerical solutions was demonstrated. In addition, the stability of networks to input data contamination (the model of channel/sensor cut-off) was estimated; the classifier’s capability to detect the dominating signal in the mixture of signals under a priori uncertainty was assessed.
The results allow drawing a preliminary conclusion: several digital signal processing problems can be effectively solved not only by statistical methods with the manual synthesis of algorithms but also by deep learning methods with the automatic synthesis of informative features and decisive rules. Moreover, as shown, convolutional neural networks can effectively work not only with signals having pronounced patterns but also with the realizations of narrow- or broad-band random processes.
Due to active research in these areas and proper work organization, the Laboratory successfully implemented several R & D projects in a short time. For example, a highly efficient primary signal classifier was developed for the oil pipeline protection system with a quantum coherent fiber-optic reflectometer (in the interests of PJSC Transneft). The signal classifier is based on a deep neural network. This project has the following characteristics: 3 real geographically separated pipelines; 6 classes of distinguishable signals (plus the background signal); 14 months of active experiments to build a library of signals and interferences; 56 interference subclasses; 70 TB of raw data. The integral F1 measure of the classifier exceeds 0.91 on the test data. A key success factor was the effective synthesis of a fast adaptive prefiltering algorithm for interferences and noises. This algorithm operates according to the blind filter principle. As a result, a classifier with high tactical and technical characteristics was developed. It operates in the hard real-time mode on common mid-priced GPUs (e.g., NVIDIA GeForce 1070).
At present, Laboratory’s employees conduct initiative research in the following areas:
- Studying the dichotomous subpartition of the TQ-space to estimate the topological entropy of discrete dynamic systems and control chaotic oscillations.
- Developing and examining constructive neural network approaches to nonlinear inverse problems.
- Designing deep neural networks of special architecture with the stabilized error of the second kind.
- Developing analysis methods for trained deep neural networks to extract learned knowledge from them.
- Developing constructive methods for training deep neural networks on quantum computers.
Thus, Laboratory No. 77 comprehensively responds to contemporary challenges in the intellectualization of control systems and processes.