Algorithm 1 Guard-Layer Action Filtering

1:Input: state ${\mathbf{s}}_t$, policy action $a_t$, threshold ${\tau }_{\mathrm{batt}}$, relaxation ${\rho }_t$   2:Output: safe action $a_{\mathrm{safe}}$   3:${\mathcal{E}}_t\leftarrow$ GetEligibleNodes$\left({\mathbf{s}}_t\right)$   4:if ${\mathcal{E}}_t=\varnothing$ then   5:    return no-transmit   6:end if   7:$i^{*}\leftarrow$Decode(at) {extract node index}   8:${\tau }^{*}\leftarrow {\rho }_t·{\tau }_{\mathrm{batt}}$ {relaxed threshold}   9:if $i^{*}\in {\mathcal{E}}_t$ and $E_{i^{*},t}>{\tau }^{*}$ then 10:    return $a_t$ {action is safe} 11:end if 12:for each $i\in {\mathcal{E}}_t$ do 13:    ${\mathrm{score}}_i\leftarrow (E_{i,t}-{\tau }^{*})+\alpha ({\overline{E}}_t-E_{i,t})$ 14:end for 15:return Encode$\left(\arg {\max }_i{\mathrm{score}}_i\right)$