[tune] Recover experiments from last checkpoint (#1532)

python/ray/rllib/agent.py

@@ -139,12 +139,19 @@ class _MockAgent(Agent):
     """Mock agent for use in tests"""
 
     _agent_name = "MockAgent"
-    _default_config = {}
+    _default_config = {
+        "mock_error": False,
+        "persistent_error": False,
+    }
 
     def _init(self):
         self.info = None
+        self.restored = False
 
     def _train(self):
+        if self.config["mock_error"] and self.iteration == 1 \
+                and (self.config["persistent_error"] or not self.restored):
+            raise Exception("mock error")
         return TrainingResult(
             episode_reward_mean=10, episode_len_mean=10,
             timesteps_this_iter=10, info={})
@@ -159,6 +166,7 @@ class _MockAgent(Agent):
         with open(checkpoint_path, 'rb') as f:
             info = pickle.load(f)
         self.info = info
+        self.restored = True
 
     def set_info(self, info):
         self.info = info

python/ray/rllib/dqn/dqn_evaluator.py

@@ -174,7 +174,9 @@ class DQNEvaluator(TFMultiGPUSupport):
             self.episode_rewards,
             self.episode_lengths,
             self.saved_mean_reward,
-            self.obs]
+            self.obs,
+            self.global_timestep,
+            self.local_timestep]
 
     def restore(self, data):
         self.exploration = data[0]
@@ -182,3 +184,5 @@ class DQNEvaluator(TFMultiGPUSupport):
         self.episode_lengths = data[2]
         self.saved_mean_reward = data[3]
         self.obs = data[4]
+        self.global_timestep = data[5]
+        self.local_timestep = data[6]