Goal-Oriented Communications for Remote Inference with Two-Way Delay

We study the design of goal-oriented communication strategies for remote inference, where an inferrer (e.g., a trained neural network) on the receiver side predicts a time-varying target signal (e.g., the position of the car in front) using the data packet (e.g., video clip) most recently received from a sensor (e.g., camera). The communication between the sensor and the receiver is carried out over a two-way channel. The objective is to minimize the expected inference error per time slot by exploiting the memory of the delay in the channel. It turns out that the optimal policy is an index-based threshold policy. The scheduler submits a packet at suitable time slots for which the index function exceeds a threshold. The index function depends on the current age of information on the receiver side and the prior knowledge about the delay in the subsequent packet transmission.