AI-Assisted MAC for Reconfigurable Intelligent Surface-Aided Wireless Networks:
Challenges and Opportunities

In the proposed centralized AI-assisted MAC framework, the BS tightly coordinates the multiple access of users. Explicitly, the BS enables each of the RIS-controllers to beneficially configure the wireless propagation environment via deep learning for multiple users. Here, the RISs are assumed to be passive, simply reflecting the incident signals without sensing or processing. More explicitly, the BS has to estimate the concatenated BS-RIS-user link, calculate the RIS phase reconfiguration, and allocate resources via deep learning, as it will be detailed below in the ‘computation’ part. As shown in Fig. 3, each sub-frame is divided into three periods: the pilot period, computing period and scheduled transmission period. The pilot period and scheduled transmission period can be further divided into $K$ pilot slots and $J$ data slots, respectively. The users transmit their data in the $J$ data slots over the $N$ non-overlapping sub-channels. Based on this, the centralized AI-assisted MAC protocol is designed by giving full cognizance to both channel estimation as well as phase computation and data transmission.

The protocol of the centralized AI-assisted MAC is shown in Fig. 3, which is integrated with TDMA and FDMA, where each user obeys the time division scheme in each sub-channel. In particular, after synchronization, each user initiates pilot transmission to the BS in dedicated pilot slots. During the computational period, the BS first estimates the concatenated RIS link, followed by time, frequency and power resources allocation and RIS phase configuration. Then the BS instructs each RIS-controller to configure its reflection parameters and schedules the access of users, who transmit their data to the BS via the RISs.

Given the excessive complexity of high-dimensional full-search-based centralized MAC protocols, a deep learning-based computational model trained offline can be employed at the BS for finding a near-optimal solution at a reduced complexity. The input of the trained deep learning model can be the number of users, the number of sub-frames, the number of channels, the number of RISs and the RIS channel information. The online inference that moves the complexity to offline training is performed at the BS, which relies on a model-based training for determining the RIS phase-shift configuration, the RIS-deployment strategy and the resource allocation strategy, which are related to each other. More explicitly, these related learning tasks share the same input parameters, thus learning multiple related tasks jointly improves the prediction accuracy and generalization capability compared to learning them separately.

Given the centralized implementation and deep learning-based computation model considered, the centralized AI-assisted MAC design advocated can be readily applied to the scenarios S1 and S3 for supporting low-power RIS-aided communications.

In contrast to the centralized scheme, in the proposed distributed AI-assisted MAC framework, each user configures the multiple access and computes the RIS configuration by itself based on the RIS-aided network environment. In this case, no BS assistance is necessary. Each RIS is assumed to be passive and to occupy a non-overlapping frequency channel. In contrast to the centralized MAC design, the user has to negotiate with the RIS-controller for channel access. The corresponding RIS-aided data transmission as illustrated in Fig. 3. In particular, channel sensing and computation are carried out at the user side via reinforcement learning (RL) to determine the RIS configuration. Once a channel is idle, the user sends the RIS configuration information to the RIS-controller to negotiate the ensuing RIS-aided data transmission.

The protocol of the proposed distributed AI-assisted MAC is shown in Fig. 3, which is integrated with CSMA and FDMA, where each user follows the distributed coordination function (DCF) based scheme in each channel. In particular, a competing user senses the state of each sub-channel. Once a channel is sensed to be idle, the user contends for the access to the channel. Waiting for a DCF inter-frame space (DISF) and backoff, the user computes its RIS configuration based on RL and sends an RIS configuration request to the RIS-controller. If the RIS is available for the user, the RIS-controller configures its reflection parameters and sends its feedback to the user after a short inter-frame spacing (SIFS). Following the elapse of a SIFS, the user then transmits the data to the BS via the RIS. Note that the feedback from the RIS-controller is sent without a transmit radio frequency chain. Moreover, the access collisions of users can be alleviated by the RIS-controller.

In the distributed MAC protocol, an RL-based computational model can be employed by each user to solve the resource allocation and RIS configuration problems because no RIS channel-information exchange is required. The RL model includes the following aspects: the current RIS configuration, the current RIS deployment, and the currently occupied resources (e.g., power) of each user. The model actions include three aspects: the update of RIS configuration, the motion-trajectory of the user and the occupied resources updates. The reward function is decided by the throughput requirement of the user. When the action taken by the user improves its data rate, the user obtains a positive reward. By contrast, for throughput reductions, the user receives a negative reward (also termed the penalty). The RL-based computational model is more suitable for small RISs to avoid potential dimensionality problems. If the RIS is large, the RIS elements may be partitioned into groups, where each group maintains the same RIS configuration.

Due to the distributed implementation and the RL-based computational model, the distributed AI-assisted MAC design can be applied to all the scenarios S1 to S4 for meeting low-latency requirements.

In this hybrid framework, the scheduled and the competing transmissions relying on RISs are combined after the pilot transmissions and computing, while enabling users to switch between them for meeting different QoS demands, as illustrated in Fig. 3. Each frame is partitioned into four parts, namely, the pilot period, the computing period, the scheduled transmission period and the competing transmission period. According to the proposed distributed AI-assisted MAC design, the scheduled users transmit their data to the BS via RISs in the scheduled transmission period. After that, based on the distributed AI-assisted MAC design, the unscheduled users (i.e., the unserviced users that have sent their pilots and the new requesting users) transmit their data to the BS or the Rxs via RISs during the competing transmission period. Given the dynamic switching between two transmission modes, this scheme is capable of maintaining the target-rate, and it may be suitable for all scenarios (i.e., S1 to S4).

In this hybrid framework, the competing requests and the scheduled RIS-aided transmissions are combined into a single frame. Each frame consists of the competing request period, the computing period and the scheduled transmission period. The user sends a request to the BS and when a sub-channel becomes available during the competing request period, then the BS issues a feedback for acknowledgment. Based on the requests received, the BS controls the RISs and sends the scheduling information to users during the computing period. Afterwards, the scheduled users transmit their data to the BS via RISs in the scheduled transmission period. Due to the competitive access of Case 2, it can be applied in scenarios S1 and S3 for supporting RIS-aided smart homes or smart factories.

This hybrid framework is similar to Case 2, since it combines the competing requests and the reserved RIS-aided transmissions into a single frame. The slight difference is that computing in Case 3 is carried out at the user, rather than at a BS or RIS-controller. When a sub-channel becomes idle, the user occupies the channel, then computes the required resources and RIS configuration based on RL and sends a request to the RIS-controller for reserving the resources for future RIS-aided transmissions. The RIS-controller sends a feedback to the user once a request is registered and controls the RIS. When the reserved transmission period arrives, the user transmits the data to the Rx or the BS via the RIS in the reserved slots. Since RL is mainly used for complexity reduction, Case 3 can be applied in all the scenarios S1 to S4 for supporting RIS-aided periodic communications.

In practice, due to the implementation constraints, the phase shifts applied by the RIS elements belong to a discrete set. Hence, the RIS configuration may be viewed as a classification problem, which can be tackled by using deep learning. If the phase shifts can be configured continuously, e.g., for varactor-based RIS designs, the optimal configuration of the phase shifts of the RIS can be regarded as a regression problem.