The Interplay of AI and Digital Twin: Bridging the Gap between Data-Driven and Model-Driven Approaches

Albeit the bright vision of realizing a holistic representation of the physical environment and network elements over a unified digital platform, the real implementation of DT can be achieved through multiple interconnected twins. While latency and complexity will be reduced within each twin, such a distributed DT implementation introduces a new level of complexity pertinent to inter-twin coordination, where models trained over the multiple twins should be aligned to ensure accurate global inference, i.e., operations over multiple twins should be synchronized to achieve joint tasks. In this process, reliability and latency should be maintained within the required thresholds. In this context, AI algorithms can be leveraged for improved twin-twin coordination. From one perspective, AI can be leveraged in order to realize minimized end-to-end (E2E) latency. In specific, graph neural network (GNN) constitutes a potential candidate for facilitating inter-twin communication and coordination. This is motivated by the fact that wireless networks modeled through multiple DTs can be represented as graphs, incorporating the global-context of twins. GNN exploits the graph structure of wireless-enabled DTs in order to capture the nodes dependencies in inter- and intra-twin communications [5]. Through the virtual twin, GNNs can develop comprehensive insights into twins and their interactions by leveraging the dynamic nature of networks as state features, and then aggregate these states to achieve a comprehensive network understanding. Recent results demonstrated the superiority of GNN in predicting the network E2E delay [6]. This metric can be of importance to accurately measure the delay encountered at the CT-CT and CT-PT communications, and compensate for that delay. From another perspective, GNNs can be exploited for improved data communication and models synchronization among multiple CTs, where the inter-twin links can be optimized to ensure synchronized twin-twin operations, and ultimately, obtain a homogeneous global DT. Additionally, goal-oriented semantic communication has two advantages in twin-twin coordination, namely, i) the proper design of the network goals to achieve the required latency, reliability, and synchronization, and ii) the cooperative operations of multiple twin to fulfill joint network goals, yielding coordinated twins at the virtual realm [7].

As discussed, one of the attractive benefits of the DT is that it constitutes a source for close-to-real data generation, in order to compensate for the weakly measured datasets at edge devices, which are diverse in terms of quality and quantity. Nevertheless, generating all-inclusive dataset that accounts for all nodes status, various environmental events, and network scenarios requires significant time resources, and therefore, is generally performed in a multi-step process. Therefore, generative adversarial networks play an important role, by employing a generator and a discriminator in order to generate accurate synthetic data. As demonstrated in Fig. 2, a generator is utilized at the CT for the purpose of generating synthetic data. Then, a discriminator is used in order to train the generator to produce higher-quality datasets, i.e., close to the data sensed from the PT. Accordingly, the real data acquired from the PT is considered as a benchmark to quantify the accuracy of the generated synthetic data from the generator.

Several variants of the GAN can introduce a different level of data generation options, and hence, offer an engineered data generation campaign, that is aligned with the data generation process at the DT. For example, conditional GAN, which impose particular constraints on the data generated by the GAN can be leveraged in order to generate datasets that are conditioned by a particular data distribution or different modality [8]. Furthermore, time-series GAN can model time-series data [9], and therefore, can assist with understanding the network dynamics over the CT. Note that GANs, and their variations, have a number of promising applications in wireless networks, e.g., channel estimation/modeling, modulation classification and recognition, and spectrum management. Accordingly, GANs represent an essential part in DT-enabled wireless networks, where synthetic data generated by GANs are expected to substantially enhance the performance of corresponding wireless networks. Within this regard, GANs at the cyber twin are anticipated to generate data with different modalities, according to the need of the considered scenario, and hence, images, RF data, etc, are potential outputs of GANs at the DT. It is worthy to note that data generated using GANs can be used to validate models trained over the DT data, and vice versa.

As discussed earlier, due to the increased complexity associated with future wireless networks, distributed DTs might be the solution for enhanced reliability and reduced latency in the DT paradigm. While on-demand data sensing is one of the operational DT pillars, it might be challenging and time-consuming to perform models retraining as a response for any environmental or network variations. This is particularly pronounced in sudden and fast variations that are generally expected in vehicular networks. Accordingly, transfer learning can be exploited to propagate the network updates over multiple twins, yielding reduced latency and overhead. Transfer learning can be applied to multiple DTs scenario, where retrained models, due to network variations, over a particular distributed twin, can be communicated with other twins with the aim to reduce the time and computing resources needed for the new models training [10]. This will further enable a more generalized distributed DT frameworks, where the knowledge in a single twin, can be applicable to other twins. While the advantages of transfer learning are particularly appealing in models retraining, it can be exploited at the initial training process of multiple twins, especially when multiple twins are experiencing noticeable variations in the measured and generated data sizes and quality. In this setup, models trained at twins with sufficiently available datasets can be communicated with other twins with the purpose of speeding up the training process at target twins. In specific, a CT with limited dataset (due to unreliable physical-cyber communication) can benefit from a well-trained model from another CT. Assuming related tasks between two CTs, the knowledge obtained from the first task training, which is based on a large and high-quality dataset, can be generalized to the second task at the other twin.

Note that transfer learning can be integrated with Reinforcement Learning (RL) for improved CT design. RL can be exploited for improved knowledge transfer between multiple CTs, where the transferred model from a well-established CT can be exploited at policy-agnostic agents in another CT, which experiences low-quality dataset, hence, rewards and actions will rely on the transferred model, and thereby, the transferred models’ weights can be fine-tuned to optimally fit the target twin. On the other hand, transfer learning can be used for policy transfer among multiple cyber twins, for fast convergence and agents training at the target CT.

As RL was initially developed as a step toward realizing autonomous systems, it constitutes a natural choice for DT applications. The merit of RL is manifested in training DTs, where a training DT can be employed for risk-controlled agents training, i.e., RL agents can freely interact with the CT, experiencing a wide range of common and rare scenarios. While maintaining the real environment unharmed is considered a big advantage, RL agents can further benefit from the DT for fast training purposes, where devices with super-computing capabilities can ensure accurately trained agents at the CT within a short period of time. These benefits are further demonstrated through a case study by Ericsson, in which they developed a DT framework to minimize the transmission power, while maintaining a particular quality-of-service (QoS) requirement and a monitored level of radio frequency (RF) radiations [11]. Allowing the RL agents to learn through direct interaction with the physical environment is considerably risky, particularly in areas with strict regulations on transmission power, and therefore, the DT offers a safe, yet efficient, virtual agents training. The optimum goal of such scenarios is to design multiple agents that are well-trained in a way that enables them to perform efficiently without further interactions with the physical environment, or to require few interactions with the real environment.

Although the recent the advancement in sensing services has facilitated data measurements campaigns, for improved AI models training process, the envisioned native AI networks will necessitate on-demand node-level data collection. This is primarily aimed in order to enable pervasive intelligence and accurate inference regardless of the network status. Such a bright vision can only be achieved if data collected represents all possible network scenarios, taking into account the physical environment status. Therefore, the DT can be a game-changer in this situation, where rarely-experienced network scenarios can be artificially engineered at the CT in order to study the network behaviour under such circumstances, and hence, perform comprehensive data collection process, that is capable of representing all nodes activities under a wide-range of network scenarios. While this data is artificially generated, it is considered close-to-real data, given that it will be generated under realistic virtual environments, that accurately imitate the dynamics of real networks. Within the same context, recalling that future wireless networks are characterized by their high level of heterogeneity, it is highly probable that local datasets at edge devices are non-identically distributed and differ in quantity and quality. This in consequence will result in models uncertainties, and hence, severely impact the network performance. \Acdt in this context ensures that models are trained over highly-reliable data in terms of quality and quantity. Note that, in order to guarantee a general-enough and accurate models when machine learning (ML) algorithms are executed in a supervised fashion, data used for testing should be different than the data used for training, however, both should be drawn from similar distributions [12]. This further corroborates the role of DT in empowering highly reliable and efficient ML algorithms, where not only large datasets can be generated, but their distributions can be controlled to ensure the required QoS.

Several research activities were initiated to explore the merit of AI when implemented in a distributed fashion. These activities were fueled by the increased communication overhead and latency, and compromised privacy, which are generally experienced in centralized methods. Although the research on distributed AI has picked up the pace in the recent years, it is still uncertain whether edge nodes are qualified for delivering the required QoS, particularly with the emergence of native AI concept, where each network node is anticipated to perform sensing, training, and inference at some level. In this regard, the DT offers a robust platform to calibrate the virtue of distributed and centralized algorithms, and to overcome their limitations. On the one hand, the overhead resulted from datasets exchange between the edge devices and the centralized aggregator will be alleviated from edge devices, users’ privacy will be maintained, and high latency encountered in centralized algorithms will be reduced. On the other hand, low training accuracy in distributed schemes, due to the local datasets limitations, can be significantly improved through the implementation of the DT paradigm, where all-inclusive datasets are available for enhanced models training. Distributed models trained at the CT might require necessary updates once implemented at the physical environment, which can be done at the edge devices at the PT, and hence, the role of edge devices will be limited to updating the local models according to the new circumstances at the physical environment. Meanwhile, in the event of operational interactive twin, models updates can be performed at the DT as well, yielding a latency-accuracy-energy trade-off problem.

The role of the mathematical frameworks in the design and optimization of wireless networks cannot be completely ignored, and a full reliance on ML tools will leave a noticeable gap. It is worthy to note that mathematical frameworks constitute solid pillars that pave the way for ML algorithms to achieve enhanced performance and scalability [15]. Therefore, despite the advanced progression in AI models, it is an imperative fact that the absence of model-driven approaches in the design and optimization process represents a bottleneck in future wireless networks. Model-driven approaches faces two major limitations: i) In some complex network scenarios, e.g., ultra-high dense heterogeneous networks, the resulted expressions representing the system are mathematically intractable, and the associated optimization problems do not lend themselves into closed-form optimum or sub-optimum solutions, and ii) The reduced accuracy when the available mathematical framework cannot be readily scaled up to represent another network scenario. Accordingly, the DT offers a way out through integrating the benefits of model-driven and data-driven approaches in a unified process. In particular, the intractability issue of some model-driven approaches can be solved through employing artificial neural networks, which is a sub-field of ML, in a way that allows ANNs to map the network parameters into the corresponding network performance, in a numerical fashion, until the convergence to the optimum solution. Such exhaustive approach is heavy and might not be tolerated by nodes with limited resources. Thanks for the DT paradigm, such a complexity can be alleviated from the physical nodes, yielding robust models training, that enjoys a solid mathematical foundation, with reduced overhead from the PT. Note that the role of DT in this scenario is to offer a platform for low-complex integration of theoretical models with ML frameworks. On the other hand, to tackle the inaccuracy limitations associated with network scalability in model-driven approaches, conventional approaches rely on initially training ML algorithms using the available mathematical models. Then, through synthetic data generation, trained models will be refined. In this scenario, the DT ensures the availability of sufficient, accurate, and close-to-real datasets for improved models refinement, and therefore, enhanced accuracy.

It is not hard to tell that the merit of the DT as enabler for AI is confined by the accuracy and the synchronization between the CT and PT. As discussed earlier, the benefits of the DT are revolving around acquiring an exact replica of the physical objects, and the instantaneous dynamics of the environment and network parameters, over a long period of time. This is due to the need of exploiting the DT as a digital environment for data generation and models training, and therefore, its accuracy is highly dependent on how accurate and realistic the CT compared to the PT. This opens the doors to further investigate the limitations of a holistic network virtualization, and to identify potential key solutions.

While the availability of large datasets at the CT for the purpose of AI models training is appealing, it is debatable whether current AI architectures will perform in a reliable and timely manner. Recalling that ultra-low latency is one of the DT verticals, it is essential to ensure that employed AI algorithms will be able deliver the required accuracy at a time-frame that is tolerable by the DT paradigm. The same applies to models updates at the CT. Hence, is there a need for designing new AI algorithms for the implementation of efficient DT? If not, to what extent current AI algorithms will be able to meet the latency requirements of the DT? This will further require the exploration of data management, cleaning, and rectifying methods, in order to enable efficient execution of AI algorithms.

Although the benefits of AI algorithms are particularly pronounced in highly dynamic, large-scale twins, such algorithms are generally treated as black boxes, where the tangible insights are obtained from the input and output data, while the complex operations in between, which are generally characterized by their extremely high complexity, remain difficult to understand, modify, and improve. In this regard, the proper understanding of the hidden operations in deep neural networks paves the way for improved architectures that better serve the needs of DT. As a promising approach, random matrix theory (RMT) has manifested itself as a versatile, yet solid, approach for analyzing DNNs. Note that, coupled with the large non-identically distributed data generated from the CT, through randomly initializing the parameters of a neural network, complex DNNs can be represented as a large matrix of random variables. Accordingly, RMT can be leveraged to optimize the initial weights of DNNs, particularly when implemented on large-scale. In specific, RMT can be efficiently exploited to design novel activation functions, that will play a role in speeding up the training process, and therefore, realize low-latency DT [15].

The proliferation of machine-to-machine communications have stimulated the recent interest in contextual-based decision-making, i.e., the so-called emergent intelligence, where multiple agents collaborate to perform a particular task through intensively interacting with each other and with the environment. While wireless networks offer an environment of a large number of communicating agents, and hence, enable the agents to develop a more advanced language that is capable of conveying a variety of abstract ideas, such scalability feature will result in either an increased latency or reduced reliability due to the substantially increased communication traffic. Within this context, DT represents an optimum candidate to enable reliable agents interaction and to support the deployment of emergent communication in future wireless networks. In particular, digital replicas of the communicating agents can be implemented at the CT, yielding improved agents information exchange, reduced latency, enhanced links reliability, reduced signaling overhead, while ensuring accurate inference [1]. As DT-empowered emergent intelligence is still in its infancy, it is essential to understand how DT can leverage the semantic aspect of communicated messages to develop a common level of understanding between communicating digital agents, that enables them to perform assigned tasks successfully, and to reflect this learning and inference process to the physical agents.