13import torch
14from torch import nn
15
16from labml_helpers.module import Module
We are using a dueling network to calculate Q-values. Intuition behind dueling network architecture is that in most states the action doesn’t matter, and in some states the action is significant. Dueling network allows this to be represented very well.
So we create two networks for $V$ and $A$ and get $Q$ from them. We share the initial layers of the $V$ and $A$ networks.
19class Model(Module):
50 def __init__(self):
51 super().__init__()
52 self.conv = nn.Sequential(
The first convolution layer takes a $84\times84$ frame and produces a $20\times20$ frame
55 nn.Conv2d(in_channels=4, out_channels=32, kernel_size=8, stride=4),
56 nn.ReLU(),
The second convolution layer takes a $20\times20$ frame and produces a $9\times9$ frame
60 nn.Conv2d(in_channels=32, out_channels=64, kernel_size=4, stride=2),
61 nn.ReLU(),
The third convolution layer takes a $9\times9$ frame and produces a $7\times7$ frame
65 nn.Conv2d(in_channels=64, out_channels=64, kernel_size=3, stride=1),
66 nn.ReLU(),
67 )
A fully connected layer takes the flattened frame from third convolution layer, and outputs $512$ features
72 self.lin = nn.Linear(in_features=7 * 7 * 64, out_features=512)
73 self.activation = nn.ReLU()
This head gives the state value $V$
76 self.state_value = nn.Sequential(
77 nn.Linear(in_features=512, out_features=256),
78 nn.ReLU(),
79 nn.Linear(in_features=256, out_features=1),
80 )
This head gives the action value $A$
82 self.action_value = nn.Sequential(
83 nn.Linear(in_features=512, out_features=256),
84 nn.ReLU(),
85 nn.Linear(in_features=256, out_features=4),
86 )
88 def forward(self, obs: torch.Tensor):
Convolution
90 h = self.conv(obs)
Reshape for linear layers
92 h = h.reshape((-1, 7 * 7 * 64))
Linear layer
95 h = self.activation(self.lin(h))
$A$
98 action_value = self.action_value(h)
$V$
100 state_value = self.state_value(h)
$A(s, a) - \frac{1}{|\mathcal{A}|} \sum_{a’ \in \mathcal{A}} A(s, a’)$
103 action_score_centered = action_value - action_value.mean(dim=-1, keepdim=True)
$Q(s, a) =V(s) + \Big(A(s, a) - \frac{1}{|\mathcal{A}|} \sum_{a’ \in \mathcal{A}} A(s, a’)\Big)$
105 q = state_value + action_score_centered
106
107 return q