Skip to main content

Actively Enhanced Cognition based Framework for Design of Complex Systems

Periodic Reporting for period 2 - AGNOSTIC (Actively Enhanced Cognition based Framework for Design of Complex Systems)

Reporting period: 2019-04-01 to 2020-09-30

Summary of the project:
Parameterized mathematical models have been central to the understanding and design of communication, networking, and radar systems. However, they often lack the ability to model intricate interactions innate in complex systems. On the other hand, data-driven approaches do not need explicit mathematical models for data generation and have a wider applicability at the cost of flexibility. These approaches need labelled data, representing all the facets of the system interaction with the environment. With the aforementioned systems becoming increasingly complex with intricate interactions and operating in dynamic environments, the number of system configurations can be rather large leading to paucity of labelled data. Thus, there are emerging networks of systems of critical importance whose cognition is not effectively covered by traditional approaches. AGNOSTIC uses the process of exploration through system probing and exploitation of observed data in an iterative manner drawing upon traditional model-based approaches and data-driven discriminative learning to enhance functionality, performance, and robustness through the notion of active cognition. AGNOSTIC clearly departs from a passive assimilation of data and aims to formalize the exploitation/exploration framework in dynamic environments. The development of this framework in three applications areas is central to AGNOSTIC. The project aims to provide active cognition in radar to learn the environment and other active systems to ensure situational awareness and coexistence; to apply active probing in radio access networks to infer network behaviour towards spectrum sharing and self-configuration; and to learn and adapt to user demand for content distribution in caching networks, drastically improving network efficiency. Although these cognitive systems interact with the environment in very different ways, sufficient abstraction allows cross-fertilization of insights and approaches motivating the joint treatment in this proposal.


O1: To devise a framework and methodology, specify the component blocks and formalize the joint design and interaction thereof

O2: To enhance situational awareness and coexistence of cognitive radar by integrated and iterative dynamic radar waveform adaptation enabled by unsupervised target discrimination within cognitive radar.

O3: To optimize cognitive RAN by facilitating autonomous deployment and optimization, along with self-organization and coexistence of future radio access networks

O4: To efficiently deliver content through active caching networks by providing seamless delivery of content for non-linear consumption over heterogeneous networks through efficient broad/multicasting and caching of popular content at the network edges taking the dynamic spatio-temporal user demand into account.

O5: To implement and experimentally validate key aspects of the active cognitive engine

Contributions and Impact:
As an alternative to conventional iterative optimization algorithms, a learning-optimization framework has been developed that can enable a fast, feasible, near-optimal solution for real-time applications. In addition, the developed learning-optimization approaches are expected to have better trade-off between computational efficiency and solution quality than conventional optimization algorithms and simple machine/deep-learning techniques.. Applying deep learning techniques to solve complicated optimization problems is still in its infancy stage. The outcome of this research will shed light on the fundamental aspects of joint design learning-and-optimization paradigm.

The research on graph representation proposed a new framework for assessing the similarity between aligned graphs based on the concept of graph spectral transformation of the Laplacian matrix. The proposed techniques showed promising results in terms of accuracy and were shown to be much less computationally intensive that previous time-domain state-of-the-art techniques. The investigation for graph representations show a better balance between complexity and performance and opens avenues to expand the investigations.

Related to graphs, the first dual networks distributed optimization algorithm with provably linear convergence was developed. This algorithm has advantages over primal distributed algorithms in that less information has to be distributed between nodes in each iteration, and in many cases the practical performance is better than those of primal algorithms.

To tackle the UAV problem, the group proposes a new architecture for UAV clustering to enable efficient multi-modal multi-task offloading. Here, the computing, caching, and communication resources are collaboratively optimized using AI-based decision making. Simulation results show that the proposed scheme significantly improves the task performance and resource efficiency of UAV clusters.

In the context of RAN optimization, a hybrid model-based and DL methods that are able to handle uncertainty due to delayed channel state information at when optimizing link parameters were developed. This hybrid approach does not incur a loss of optimality. The project has also extended Bayesian methods from reinforcement to the problem of link level adaptation with target error-rate constraints. This enables a fully cognitive approach where the transmitted is able to optimally trade of exploration and exploitation in the data transmission.

Concerning radar system design, an efficient optimization of the transmit waveform, receiver algorithms and radar architecture is considered, These building blocks, when further augmented by additional information about the environment and the on-going waveform adaptation mechanisms, will lead to the successful demonstration of the AGNOSTIC framework in cognitive radar.
AGNOSTIC envisions a computationally efficient framework for enhancing the performance of complex dynamical cognitive systems by actively probing the environment/system and allowing model-based and data-driven approaches to work in tandem to overcome their respective shortcomings. Five work packages have been envisioned to achieve this objective. The work pursued during the reporting period (corresponding to about 18 months of the project) include:

1) (WP1) Investigations to devise a framework for formalizing the active learning using data-driven and model-based approaches which would result in the AGNOSTIC framework. In addition to tackling the problem from this objective, a top-down approach of investigation how this aspect has been dealt in the various AGNOSTIC application areas has been pursued. Mathematical approaches including iterative optimization for large data sets are being considered.

2) (WP2-4) Initial investigations into the use of data-driven and model-based system design for radar, communication and caching in static environments have been initiated. This study is a precursor to the active learning for dynamic systems. Results from WP2-4 will be analyzed and considered for in-lab demonstration in WP5 (to start in year 4).

The project has resulted in 24 conference articles, 18 journal papers and two book chapters till date. The work pursued in these publications are summarized in the other sections of this form appropriately.

The topic of AGNOSTIC was selected for a special session at the prestigious IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, Calgary, Canada ( ). It provided visibility to the research and also a platform for discussions. To foster interaction between the involved parties, a joint workshop was held in October 2018, followed by planned research visits. An invitation to deliver a plenary keynote presentation at IEEE ICASSP 2020 has been accepted by the main PI, Björn Ottersten. Further, Dr Bhavani Shankar MYSORE's proposal with colleagues to deliver a tutorial at IEEE ICASSP 2020 and IEEE RadarConf 2020 were also accepted.
Work Package 1:

Most of the ML\DL works in wireless applications are based on a static design, i.e. based on a collected data set, a snapshot of a deterministic complex network, which is used to develop optimization solutions. In addition, the inputs, e.g. channel, traffic, for training an ML/DL model are typically required to follow a certain probability distribution. However, a realistic complex network is often intertwined with dynamicity, e.g. fast variations of network topology, bursty traffic, terminals entry/leave, dramatic changes of channel conditions. In practice, the after-learned model might be valid for a short time, but could become invalid when the inputs no long follow the distribution. In the worst case, the ML/DL model can significantly degrade the performance, and have to be re-trained along with regenerating/recollecting new data sets. In such cease, it is essential and also challenging to consider how to avoid re-training an ML/DL model by long time, and make it adaptive to dynamic environment timely. Towards this end, we extend the developed data-driven framework to deal with such network dynamicity. Several DL-based tailored solutions, e.g. transfer learning, meta learning, deep reinforcement learning, are being developed for addressing various dynamic scenarios.

Novel Graph Signal Processing (GSP) methodologies for finding anomalies in network signals using vertex-frequency analysis to extract features that feed machine learning algorithms has been developed. In addition, a normalized least-mean-squares (NLMS) algorithm for online estimation of bandlimited graph signals using a reduced number of noisy measurements is proposed. The current investigation is on extending the obtained results to deal with kernel-based distributed adaptive graph filters, as well as forecasting and interpolation of graph signals using deep learning algorithms. The expected solutions will be able to learn highly complex non-linear mappings of network signals in an online manner, as well as to solve the problem of spatiotemporal prediction of network signals with highly non-linear space-time dynamics.

Work Package 2:

Towards enhancing the information extraction capabilities in a radar receiver- the key to enhanced cognition, investigation of novel estimators for the co-array-based Direction of Arrival estimation is pursued. The parameters of the estimator are then optimized to ensure asymptotic statistical efficiency of resulting estimator. Further, in the context of sparse frequency diverse array radars, waveform and array locations have been optimized. In addition to the radar received data, other sources have also been considered to offer side-information. In this context, investigation on joint radar-communications as a means to augment cognition in automotive systems is underway,

Work Package 3:

Research progressing beyond state of the art has continually been published in peer-reviews scientific journals and conference, as shown by the list of publications. Work at KTH with WP3 will for the remainder of the project focus on Link Adaptation and Power Control, and Hybrid Approaches. In particular, we are expecting to obtain novel results from the merger of convex optimization algorithms and the computational frameworks developed for deep learning, though so called deep unfolding techniques applied within the communications (RAN) context.

Although ML techniques have been recently applied to enhance the cooperative sensing performance in CRNs, they are mostly supervised learning-based techniques and need a significant amount of labeled data, which is difficult to acquire in practice. Towards relaxing this requirement of large labeled data of supervised learning, we have focused on Active Learning (AL), where the fusion center can query the label of the most uncertain cooperative sensing measurements. This is particularly relevant in CRN environments where primary user behavior changes in a quick manner. Results based on numerical simulations show that the proposed method has significant advantages on classification and detection performances, and time-complexity as compared to state-of-the-art techniques.

Work Package 4:

The developed caching strategies requires is partially model-based which requires the network topology as well as the popularity model to be static. In practice, users are usually mobile and therefore joint and leave a service in a random manner. As a result, the system model and network topology can significantly change over time, which requires novel content delivery methods adapted to the system changes. The current supervised learning-based solutions might be inappropriate in such cases because there is not sufficient data to train the learning model. Our plan is to exploit deep learning and reinforcement learning tools to predict the temporary content popularity in such highly dynamic environment. While deep learning is powerful to learn inherent property/feature of the system, reinforcement learning offers fast-response decision and does not require much training data. The expected solutions will be able to learn the system dynamics and properly estimate both local and global content popularities.