Automatic AI Model Selection for Wireless Systems: Online Learning via Digital Twinning

TL;DR

A novel method for the online optimization of AMS mapping that corrects for the bias of the simulator by means of limited real data collected from the physical system is proposed.

Abstract

In modern wireless network architectures, such as O-RAN, artificial intelligence (AI)-based applications are deployed at intelligent controllers to carry out functionalities like scheduling or power control. The AI "apps" are selected on the basis of contextual information such as network conditions, topology, traffic statistics, and design goals. The mapping between context and AI model parameters is ideally done in a zero-shot fashion via an automatic model selection (AMS) mapping that leverages only contextual information without requiring any current data. This paper introduces a general methodology for the online optimization of AMS mappings. Optimizing an AMS mapping is challenging, as it requires exposure to data collected from many different contexts. Therefore, if carried out online, this initial optimization phase would be extremely time consuming. A possible solution is to leverage a digital twin of the physical system to generate synthetic data from multiple simulated contexts. However, given that the simulator at the digital twin is imperfect, a direct use of simulated data for the optimization of the AMS mapping would yield poor performance when tested in the real system. This paper proposes a novel method for the online optimization of AMS mapping that corrects for the bias of the simulator by means of limited real data collected from the physical system. Experimental results for a graph neural network-based power control app demonstrate the significant advantages of the proposed approach.

Figures (11)

• Figure 1: As a use case of the proposed methodology, this figure illustrates the O-RAN architecture, in which a base station (BS), known as gNB, is disaggregated into a central unit (CU), a distributed unit (DU), and a radio unit (RU), where both CU and DU are deployed in the O-Cloud, and the RU is deployed at the BS. The near-real time (RT) radio intelligent controller (RIC) deploys AI-based xApps to carry out functionalities at different layers of the protocol stack. As shown in part (a), in the considered setting, an automatic model selection (AMS) mapping produces the parameters $\phi$ of an AI model based on context information from the BS. The AI model may be implemented at the near-RT RIC for an O-RAN system. The focus of this work is on the online optimization of the AMS mapping, which is carried out during a preliminary calibration phase. Specifically, as shown in part (b) for the O-RAN architecture, this work proposes to speed up online calibration via the use of a digital twin, which can produce synthetic data for new contexts. In the proposed scheme, for each real-world context $c$, the digital twin produces synthetic data from multiple simulated contexts, and the real-world data is leveraged to "rectify" the errors made by the simulator.
• Figure 2: Wireless resource allocation using AMS: Given the context $c$ describing the matrix of geographical distances among the network nodes, i.e., the network topology, the AMS mapping $\phi=g(c)$ returns a GNN model $\phi$. The GNN model takes the current network CSI as input, and it outputs the transmission powers for all the nodes.
• Figure 3: Conventional AMS calibration based on online learning: At each time step $t$, the physical twin (PT) is faced with a context $c_t\sim p(c)$, and it tests the current AMS output model $\phi_t=g(c_t|\theta_t)$ over $N^{\text{PT}}$ channel realizations, producing the estimate (\ref{['PT loss']}). Using online gradient descent, the AMS model parameter vector is updated using (\ref{['onlyPT']}).
• Figure 4: DT-powered AMS calibration based on online learning: At each time step $t$, the PT is faced with a context $c_t\sim p(c)$, and it collects corresponding real-world data $x_{n,t}\sim p(x|c_t)$ for $n=1,\dots,N^{\text{PT}}$. The PT sends the current context $c_t$ to the DT, which generates corresponding simulated data $\tilde{x}_{n,t}\sim p(\tilde{x}|c_t,f)$ for $n=1,\dots,N^{\text{DT}}$, as well as data $\tilde{x}_{m,n,t}\sim p(\tilde{x}|c,f)$ for $n=1,\dots,N^{\text{DT}}$ under $M^{\text{DT}}-1$ independent context variables $c_{m,t}\sim p(c)$ for $m=1,\dots,M^{\text{DT}}-1$. The DT uses the synthetic data points to evaluate the empirical losses (\ref{['DT loss']}) and (\ref{['DT loss tilde ppi']}). Upon reception of the model $\phi_t$ from the DT, the PT runs it over $N^{\text{PT}}$ channel realizations, producing the loss estimate (\ref{['PT loss']}). Using online gradient descent, the AMS model parameter vector is updated using (\ref{['ppi']}).
• Figure 5: Time-varying $K$-user interference network assumed in the experimental results.