Algorithm 2 Heterogeneous DDLO for MEC networks
1:	Input: all WDs’ task requirements ${\mathit{s}}_t$
2:	Output: offloading decision ${\mathit{a}}_t^{*}$
3:	Initialization:
4:	Initialize all B DNNs with different random parameters ${\theta }_t^b$, $b\in \mathcal{B}$;
5:	Initialize memory structure with size H;
6:	for $t=1,2,\dots ,G$ do
7:	Input the same ${\mathit{s}}_t$ to each DNN.
8:	Generate B offloading action candidates from the DNNs $\left\{{\mathit{a}}_t^b\right\}=f_{{\theta }_t^b}\left({\mathit{s}}_{\mathit{t}}\right)$;
9:	Select the offloading decision ${\mathit{a}}_t^{*}=\underset{b\in \mathcal{B}}{argmin}Q({\mathit{s}}_t,{\mathit{a}}_t^b)$;
10:	Store $({\mathit{s}}_t,{\mathit{a}}_t^{*})$ into the memory structure;
11:	Randomly Sample B batches of training data from the memory structure;
12:	Train the DNNs;
13:	end for