Advances, Systems and Applications
From: Joint DNN partitioning and task offloading in mobile edge computing via deep reinforcement learning
Notation | Definition |
---|---|
n | The number of MDs |
\(D_u(t)\) | The task size on the uth MD in time slot t |
\(Q_u(t)\) | The task size currently stored in the buffer of the uth MD in time slot t |
\(L_u\) | The amount of layers of the DNN model on the uth MD |
\(\alpha _{u,L}(t)\) | The amount of layers computed on the uth MD in time slot t |
\(\alpha _{u,M}(t)\) | The amount of layers computed on the MEC server in time slot t |
\(Q_{u}^{'}(t)\) | The unexecuted task size on the uth MD in time slot t |
A | The collection of all the action tuples |
\(M_{u,l}\) | The subtask l on the uth MD |
\(f_{u,l}\) | The ratio of input matrix of the DNN layer l on the uth MD to the initial data size \(D_u(t)\) |
\(Y_{u,l}(t)\) | The input matrix size of the subtask \(M_{u,l}\) in time slot t |
\(\xi _{u,loc}(t)\) | The processing time taken by the uth MD to process one unit of data in time slot t |
\(\xi _{u,loc}^{min}\) | The minimum processing time taken by the uth MD to process one unit of data |
\(\xi _{mec}(t)\) | The time taken by the server to process one unit of data in time slot t |
\(\xi _{mec}^{min}\) | The minimum time taken by the server to process one unit of data |
\(B_u(t)\) | The bandwidth between the uth MD and the MEC server in time slot t |
\(N_0\) | The power spectral density of noise |
\(h_u(t)\) | The channel power gain between the uth MD and the MEC server in time slot t |
\(P_{u}^{up}(t)\) | The uploading power of the uth MD in time slot t |
\(P_{M}(t)\) | The transmission power of the server in time slot t |
\(O_{k_{u}}(t)\) | The output data size from the DNN layer \(k_u\) on the uth MD in time slot t |
\(O_{L_{u}}(t)\) | The output data size from the last DNN layer \(L_{u}\) on the uth MD in time slot t |
\(P_{u}^{exe}(t)\) | The computing power of the uth MD in time slot t |
S | The collection of all the states |
\(C_u(t)\) | The overall cost for the uth MD in time slot t |
\(R_u(t)\) | The reward value corresponding to \((s_u(t),a_u(t))\) |