Always forward state for recurrent nets

Kaixhin · Jun 26, 2016 · 626712b · 626712b
1 parent 6e66251
commit 626712b
Show file tree

Hide file tree

Showing 2 changed files with 10 additions and 0 deletions.
diff --git a/Agent.lua b/Agent.lua
@@ -193,6 +193,11 @@ function Agent:observe(reward, rawObservation, terminal)
       -- Choose action by ε-greedy exploration (even with bootstraps)
       aIndex = torch.random(1, self.m)
 
+      -- Forward state anyway if recurrent
+      if self.recurrent then
+        self.policyNet:forward(state)
+      end
+
       -- Reset saliency if action not chosen by network
       if self.saliency then
         self.saliencyMap:zero()

diff --git a/async/QAgent.lua b/async/QAgent.lua
@@ -20,6 +20,7 @@ function QAgent:_init(opt, policyNet, targetNet, theta, targetTheta, atomic, sha
   self.dTheta:zero()
 
   self.doubleQ = opt.doubleQ
+  self.recurrent = opt.recurrent
 
   self.epsilonStart = opt.epsilonStart
   self.epsilon = self.epsilonStart
@@ -60,6 +61,10 @@ function QAgent:eGreedy(state, net)
   end
 
   if torch.uniform() < self.epsilon then
+    -- Forward state anyway if recurrent
+    if self.recurrent then
+      net:forward(state)
+    end
     return torch.random(1,self.m)
   end