Algorithm 2: Actor Network Forward Pass for Joint Action Generation

$\mathbf{Input}:\mathrm{GNN}\mathrm{node}\mathrm{embeddings}\mathrm{H},\mathrm{vehicle}\mathrm{embedding}{\mathcal{h}}_{veh}$$,\mathrm{task}\mathrm{features}{\mathcal{X}}_{task}$$,\mathrm{global}\mathrm{stats}{\mathcal{g}}_{stats}$

$\mathbf{Output}:\mathrm{Joint}\mathrm{action}\left({\mathcal{K}}_{\mathcal{i}},{\lambda }_{\mathcal{i}},{\alpha }_{\mathcal{i}}\right)$$,\log \mathrm{probability}\log \pi \left({\mathcal{K}}_{\mathcal{i}},{\lambda }_{\mathcal{i}},{\alpha }_{\mathcal{i}}|{\mathcal{S}}_t\right)$

1: Phase A: Select Offloading Target via Attention

$2:{\mathcal{q}}_{att}$$\leftarrow {MLP}_{query}\left(\left[{\mathcal{h}}_{veh}\oplus {\mathcal{X}}_{task}\oplus {\mathcal{g}}_{stats}\right]\right)$

$3:C_{target}$$\leftarrow [h\_veh,h\_rsu\_1,...,h\_rsu\_M]$

$4:\mathbf{for}\mathrm{each}\mathrm{candidate}{c}_{\mathcal{j}}$$\mathrm{in}{C}_{target}$ do

$5:{\mathcal{e}}_{\mathcal{j}}$$\leftarrow ({\mathcal{q}}_{att}$$\cdot c_{\mathcal{j}}$$)/$$\mathrm{sqrt}({\mathcal{D}}_{emb}$)

6: end for

7: Mask scores for invalid/unavailable targets

$8:P_{target}$$\leftarrow Softmax$$\left(\right[{\mathcal{e}}_0,{\mathcal{e}}_1,\dots ,{\mathcal{e}}_M\left]\right)$

$9:\mathrm{Sample}\mathrm{target}\mathrm{index}{\mathcal{K}}_{\mathcal{i}}$$~\mathrm{Categorical}(P_{target}$)

$10:{\mathcal{h}}_{target}$$\leftarrow C_{target}$$[{\mathcal{K}}_{\mathcal{i}}$]

11: Phase B: Determine Continuous Ratios

$12:\mathbf{if}{\mathcal{K}}_{\mathcal{i}}$ = 0 then

$13:{\lambda }_{\mathcal{i}}$$\leftarrow 0,{\alpha }_{\mathcal{i}}$ ← 1

14: else

$15:a_{\lambda },b_{\lambda }$$\leftarrow Softplus\left({MLP}_{\lambda }\left(\left[{\mathcal{h}}_{veh}\oplus {{\mathcal{h}}_{target}\oplus \mathcal{X}}_{task}\oplus {\mathcal{g}}_{stats}\right]\right)\right)$ + 1

$16:\mathrm{Sample}\mathrm{offload}\mathrm{ratio}{\lambda }_{\mathcal{i}}$$~Beat\left(a_{\lambda },b_{\lambda }\right)$

$17:a_{\alpha },b_{\alpha }$$\leftarrow Softplus\left({MLP}_{\alpha }\left(\left[{\mathcal{h}}_{veh}\oplus {{\mathcal{h}}_{target}\oplus \mathcal{X}}_{task}\oplus {\mathcal{g}}_{stats}\oplus {\lambda }_{\mathcal{i}}\right]\right)\right)$ + 1

$18:\mathrm{Sample}\mathrm{resource}\mathrm{ratio}{\alpha }_{\mathcal{i}}$$~Beat\left(a_{\alpha },b_{\alpha }\right)$

19: end if

20: Calculate total log probability log π from the distributions

$21:\mathbf{return}\left({\mathcal{K}}_{\mathcal{i}},{\lambda }_{\mathcal{i}},{\alpha }_{\mathcal{i}}\right)$$,\log \pi \left({\mathcal{K}}_{\mathcal{i}},{\lambda }_{\mathcal{i}},{\alpha }_{\mathcal{i}}|{\mathcal{S}}_t\right)$