🚑 Fix learn failing

Make sure that default learn_start is greater than the minibatch size

arcades/agent/NeuralQLearner.lua

@@ -92,10 +92,6 @@ function class:__init(args, dump)
 
   self:_init_inference_network(args.inference, dump.inference)
 
-  --- Number of steps after which learning starts.
-  -- Add delay to populate the experience pool
-  -- @tfield number self.learn_start
-  self.learn_start = args.learn_start or (0.1 * self.experience_pool.pool.max_size)
   --- Learning frequency (epoch size).
   -- @tfield number self.update_freq
   self.update_freq = args.update_freq or 12
@@ -105,6 +101,10 @@ function class:__init(args, dump)
   --- Number of minibatch learning during a learning epoch.
   -- @tfield number self.n_replay
   self.n_replay = args.n_replay or 1
+  --- Number of steps after which learning starts.
+  -- Add delay to populate the experience pool
+  -- @tfield number self.learn_start
+  self.learn_start = args.learn_start or math.max(self.minibatch_size + 1, (0.1 * self.experience_pool.pool.max_size))
 
   --- Scale rewards delta.
   -- @tfield boolean self.rescale_r