Validate FMASAC under constraints, heterogeneity, and increasing scale

Determine whether the federated multi-agent soft actor–critic (FMASAC) framework for IRS-assisted cognitive radio networks remains effective under stricter interference-constraint handling, heterogeneous secondary-user channels and locations, and increasing numbers of secondary users and IRS elements, thereby validating the framework beyond its illustrative evaluation.

Background

The paper presents FMASAC as a framework in which distributed secondary users jointly learn transmit-power and IRS-phase decisions from local observations, while a federated server periodically averages actor and critic parameters. The reported performance curves are synthetic illustrations rather than results from trained agents in a simulated network, so the framework has not yet been empirically validated.

The authors explicitly identify constraint handling, agent heterogeneity, and scalability as open challenges. In particular, the interference constraint is enforced only through a soft reward penalty; the federated averaging rule assumes that secondary users are sufficiently similar; and the effectiveness of IRS-element partitioning and parameter averaging as the network grows remains to be tested. The evaluation plan proposes comparing penalty-based learning with primal–dual or Lagrangian SAC, testing non-identically distributed users and personalized aggregation, and varying the numbers of users and IRS elements.