Deep Learning based Channel Assignment

To better describe our proposal, we divide the whole assignment process into two steps, i.e., the training phase and dynamic channel assignment phase.

5.1.1.1 Training Phase

Here, we use the traffic load and channel assignment result of AC-POCA as the training data set. Before using the data set, we need to characterize the training data into a suitable format. As we described above, we use the traffic load as the main feature to construct the input of training data. Such training data format is denoted as tl_k ={tl_k¹,tl_k²,· · · ,tl_k^M}.

Because the AC-POCA only considers the current traffic load, and the result is not affected by the traffic load of the past time sequence, we only use the traffic load of the last one slot as the input of training data.

Then, we consider that the assigned number of each link is recorded as the output of training data. If the scale of the network is significantly large, the number of links is large to make the output very complex. As mentioned in our previous work [100], the complex output will significantly decrease the training accuracy. Therefore, as the same method employed in our traffic prediction algorithm, we useM×E_max neural network to separately train the network, where E_max denotes the maximum number of active links of each node. For example, for 802.11 2.4 GHz links, because of self-interference, the same channel cannot be assigned to two links of one node. Then, the maximum number of active links E_max is 11, which is equal to the maximum number of channels C_max. Since each neural network is only used to predict the channel for one link, the number of total neural networks is equal to the number of links. And the neural network corresponding to the j^th link in switch j is recorded as {NN_i,j|i ≤M,j ≤E_max}. For each neural network, the output is characterized as a vector consisting of C_max binary elements, which can be denoted as L={l₁,l₂,· · · ,l_Cmax−1,l_Cmax|l ∈ {0,1}}. And if channel i is assigned, the

0 500 1000 1500 2000 2500 3000 3500

0 50 100 150 200 250 300 350 400 450 500

Packets generation rate (kbps)

Time (s)

switch 1 switch 2 switch 3

Figure 5.2: The integrated traffic pattern of different switches in the periodic intensive case (PIC).

value of the i^th element is 1, otherwise 0. Therefore, the training data of each neural network is indicated as (x_input,y_output) = (tl_k,L).

With the training data, we try different kinds of neural network structures and different parameters for training. The comparison of the training results of different structures and parameters is shown in Sec. 5.2.2. Because of the large number of neural network and data set, this training process is better to be processed in the central controller. And the bias and weight matrices of all neural networks are recorded and updated in the central con-troller. The trained weight matrix of each switch is recorded as{WM_i,j|i ≤M,j ≤E_max}.

PIC EIC

R 100 100

Q 30 30

W 10 20

n_p 80 50

n_e 20 50

Table 5.1: The configuration of PIC and EIC

0 500 1000 1500 2000 2500 3000 3500

0 50 100 150 200 250 300 350 400 450 500

Packets generation rate (kbps)

Time (s)

switch 1 switch 2 switch 3

Figure 5.3: The integrated traffic pattern of different switches in the event intensive case (EIC).

5.1.1.2 Dynamic Channel Assignment Phase

After the training process, the central controller sends the copy of the trained weight matrices to each corresponding switch. Each switch only stores the weight matrices cor-responding to its own links. Then, during the packet transmission period, the central controller sends the traffic load information tlk to each switch periodically, and each switch uses the current traffic load information as the input to trigger a forward propaga-tion process with the corresponding weight matrix to get the outputL^k+1 (i.e., the binary vector of chosen channel). If the already assigned channelL^k of the link is different from the new one L^k+1, the switch confirms the new channel number with the other switch on the other side of the link. If both switches get the same result, they change the channel of this link to the new one. Otherwise, the switches report the different results to the central controller. The whole process is shown in Alg. 2.

With deep learning based channel assignment, the channel assignment result can be simply obtained via a forward propagation process, which is much faster than the the game theory based channel assignment and saves most of the communication cost/signaling overhead. This is because in the conventional game theory based channel assignment methods, each router needs to keep receiving and updating the channel statements of all other routers in every iteration. Consequently, the more iterations of decision process, the heavier signaling overhead. On the other hand, in our proposed deep learning based

Algorithm 2 Algorithm of DLPOCA

Input: Trained weight matrices {WMi,j|i≤M,j≤Emax}; each switch{si|i = 1 → M}, traffic load tlk.

Output: The assigned channelL^k+1_j of each link{ej|j= 1→Emax}.

1: forj= 1 toEmax do

2: The switch on the other side of linkej is record as{sf|f ≤M}.

3: si usetlk as input to trigger forward propagation withW Mi,j, the result is recorded asL^k+1_j . 4: if L^k+1_j 6=L^k_j then

5: if L^k+1_j =L^k+1_f ofs_f then 6: AssignL^k+1_j to linke_j.

7: else

8: Feedback the wrong information to central controller.

9: end if 10: end if 11: end for

channel assignment, only one iteration is needed, which is the main reason why our proposal can significantly outperform the conventional one.

5.1.2 Deep Learning based Channel Assignment jointed with