Fluctuation-Induced Network Control and Learning

What are the underlying principles that explain how complex biosystems work in such a remarkably energy-saving and flexible manner? In this chapter we explore this question based on our research of muscle and brain, both of which are typically complex biosystems. Our state-of-the-art imaging technology of direct observation of individual molecular motor motion has revealed the surprising fact that muscle contraction is produced by utilizing the Brownian motion of molecular motors in a very skillful manner. This indicates that the muscle can utilize thermal fluctuations of molecules in an effective way, such that the energy efficiency of muscle is extremely high compared to artificial energy conversion systems. Although the research methods are quite different, we have also observed analogous findings in human brain function. Our research on human visual recognition showed that the time taken for recognizing a difficult figure follows the same exponential function as the “Arrhenius equation,” which describes how the rate of a chemical reaction is driven by thermal fluctuation. This finding, as well as our related modeling, strongly supports the idea that stochastic activity, possibly resting-level spontaneous activity, may also help make human recognition flexible and energy saving. Based on these findings, we would argue that fluctuations in biosystems, both thermal fluctuation of motor molecules in muscle and stochastic activity of neurons in the brain, play an essential role in flexible and energy-saving functioning. To emphasize these positive aspects of fluctuations in biosystems, we would propose the concept of “Yuragi,” which is a word of Japanese origin with the meaning of fluctuations for flexible adaptation to the environment. We would suggest that the utilization of Yuragi is one of the principles for efficiency and flexibility of biosystems.

The Yuragi model can be regarded as a meta-heuristic algorithm for an optimization problem whose conditions change over time. The activity in the Yuragi equation dynamically strikes a balance between reinforcement learning, which contributes to stability, and random search, which aims to find a solution appropriate for a new condition. Network control methods must provide stable network service by continuously adapting to changing conditions. No information networks are stable; instead, they are exposed to internal and external perturbations, thereby leading the perceived quality and performance to always fluctuate. Using the Yuragi model, a system can adaptively select a state or control option that is appropriate for the current environmental condition, stably remain there as long as the condition does not significantly change, and switch to a new appropriate state once the condition changes. In this chapter, we introduce examples of applications of the Yuragi model to network control.

Reconfiguring a virtual network, which consists of a set of virtual links and routers, on top of a physical network is a promising approach to accommodate time-varying traffic. However, optimization-based approaches are generally incapable of quickly adapting to time-varying traffic due to their computational complexity. For swift adaptation to changes in traffic patterns and network failures, this chapter presents Yuragi-based virtual network control. This control is based on attractor selection, which models behavior whereby biological systems adapt to unknown changes in their surrounding environments. A biological system driven by attractor selection adapts to environmental changes, selecting attractors, in which the condition of the system is well suited for a certain environment, by using noise, also referred to as Yuragi. There are two main challenges in achieving virtual network control based on attractor selection. The first involves determining how to map the behavior of a biological system to virtual network control, while the second involves designing attractors for robust virtual network control. This chapter summarizes previous studies that solved these two research problems.

Yuragi learning is a paradigm that uses attractors as state templates of network systems. Under uncertain or fluctuating traffic in the network systems, template matching with predefined attractors is first examined, and the state of the network system is then changed to the most appropriate attractors. In this introductory chapter to Yuragi learning and its applications, we present a fundamental model of human perceptual decision-making, which is the core of Yuragi learning.

Due to power and space considerations, the link capacity in many Internet of Things (IoT) networks is low; however, many IoT sensors, such as high-resolution video cameras, generate huge amounts of data, which can cause congestion unless traffic engineering is applied. However, the majority of the existing network traffic engineering methods require traffic matrix information, which can be difficult to estimate in IoT networks. Instead of attempting to estimate the traffic matrix, we identify the traffic pattern using machine learning based on the Bayesian attractor model (BAM) for supervision and automation of traffic engineering in IoT networks that exhibit a limited number of traffic patterns. We propose running two BAMs in parallel: a fast-pathway BAM for fast but low-certainty identification, and a slow-pathway BAM for slow but high-certainty identification. We demonstrate that our framework enables fast and reliable identification of traffic patterns. After identifying a traffic pattern, a network configuration that is optimized for the identified pattern by traffic engineering is applied to minimize the maximum link utilization. In traffic engineering, we apply virtual network slicing, which creates an independent end-to-end logical network for each IoT sensor type on a shared physical infrastructure. We demonstrate that virtual network slicing allows for fine-grained traffic engineering in IoT networks.

In this chapter, I describe the application of the Yuragi learning mechanism to the control of networks that accommodate Internet of Things (IoT) services. Numerous types of services have been provided through networks as IoT devices have become increasingly popular, and the traffic resulting from such services must be accommodated to satisfy the service-specific requirements. One approach to accommodate traffic is the use of network slicing, which provides multiple network slices for network services. Resources for each slice should be dynamically allocated to follow the traffic changes. Predictive network control is a method for accommodating fluctuating traffic without degrading service quality. This method allocates resources based on predicted future traffic. To predict future traffic, real-world information is useful. Thus, in this chapter, we discuss the use of real-world information related to IoT services to predict future traffic. However, it is difficult to model the relationship between real-world information and future traffic; therefore, we apply Yuragi learning, which is inspired by the cognitive processes of the human brain and makes decisions based on uncertain information. In this chapter, we discuss the application of Yuragi learning to predictive network control based on real-world information.

The Internet of Things has become increasingly widespread, and low-power wide-area (LPWA) technology has attracted attention as one of its elemental technologies. LPWA technology achieves wide-area communication without consuming a large amount of energy, which facilitates various types of applications for sensing and collecting data. LoRa (Long Range) is a type of LPWA communication technology that uses unlicensed bands. Because it is possible to build a self-managed network with LoRa, many services using LoRa are scattered in the same area without an administrator. As a result, the communication performance of LoRa may be degraded due to unintended radio interference. However, because many LPWA techniques, including LoRa, have a low data rate, it is difficult to gather sufficient control information to avoid the degradation of communication performance. In this chapter, we propose a method for predicting the network congestion state based on Yuragi learning that enables prediction from fluctuating and noisy data by successive Bayesian estimation. Through computer simulation, we demonstrate that the network state can be predicted by our proposed method with a small amount of control information.

Neural network-based artificial intelligence, such as deep learning, has made remarkable progress by utilizing high-performance computing resources. However, research in neuroscience has revealed that the human brain can realize multimodal, flexible, and efficient cognitive operations by consuming only approximately 21 W of power. This chapter summarizes our challenges in developing an artificial cognitive system based on the Bayesian attractor model and provides a brief introduction to our software prototype called the Yuragi Learning General-Purpose Data Analysis Platform (YGAP).

One of the differences between traditional artificial intelligence (AI) and human cognition is the amount of training data required to learn a new category. Conventional AI, such as deep learning, must learn a large amount of data to make decisions; however, humans can make decisions using only several representative examples of data. This ability is essential, especially in situations in which the environment frequently changes, such as the Internet of Things (IoT). With frequent changes in the environment, AI must make decisions based on new situations in real time. Traditionally, several learning methods have been proposed, such as self-training and transfer learning, which assume a small amount of training data. However, when the environment constantly changes and real-time decisions are required, it is necessary to learn and make decisions with a small amount of available data. For this purpose, Yuragi learning, which is a model of decision-making that assumes that obtained information contains uncertainty, is promising due to its noise tolerance. In this chapter, we introduce a system that uses Yuragi learning as a classifier to automatically acquire a new category in new situations. This system detects new categories based on the confidence of existing categories computed by Yuragi learning. Then, the detected categories are added, and new training data are collected in parallel with ordinal classification. The collected data are used as training data for the new category, and additional training is performed to improve the classification accuracy. In this chapter, we apply this classification system to an image classification task for handwritten characters as an example of its application.