comments

ReinforceGamma.lua

@@ -13,8 +13,8 @@
 local ReinforceGamma, parent = torch.class("nn.ReinforceGamma", "nn.Reinforce")
 
 function ReinforceGamma:__init(scale, stochastic)
-   require('randomkit') -- needed to sample gamma dist
-   require('cephes') -- needed to compute digamma for gradient
+   require('randomkit') -- needed to sample gamma dist : luarocks install randomkit
+   require('cephes') -- needed to compute digamma for gradient : 
    parent.__init(self, stochastic)
    self.scale = scale
    if not scale then

test/test.lua

@@ -1669,13 +1669,15 @@ function dpnnbigtest.Reinforce()
             end
             reward = reward/inputs:size(1)
 
-            if reward*0.7 >= baseReward then
+            -- is the baseReward lesser than 70% of reward after training?
+            -- i.e. did the reward increase sufficiently?
+            if reward*0.7 > baseReward then
                converged = true
                break
             end
          end
 
-         if reward*0.7 >= baseReward then
+         if reward*0.7 > baseReward then
             converged = true
             break
          end