[rllib] Add n-step Q learning for DQN (#1439)

* n-step

* add sample adjustm

* Oops

* fix nstep

* metric adjustment

* Sat Jan 20 23:30:34 PST 2018

* Sun Jan 21 16:40:46 PST 2018

* Mon Jan 22 22:24:57 PST 2018

python/ray/tune/trial.py

@@ -42,6 +42,9 @@ class Resources(
         return super(Resources, cls).__new__(
             cls, cpu, gpu, driver_cpu_limit, driver_gpu_limit)
 
+    def summary_string(self):
+        return "{} CPUs, {} GPUs".format(self.cpu, self.gpu)
+
 
 class Trial(object):
     """A trial object holds the state for one model training run.

python/ray/tune/trial_runner.py

@@ -79,9 +79,12 @@ class TrialRunner(object):
             for trial in self._trials:
                 if trial.status == Trial.PENDING:
                     if not self.has_resources(trial.resources):
-                        raise TuneError(
-                            "Insufficient cluster resources to launch trial",
-                            (trial.resources, self._avail_resources))
+                        raise TuneError((
+                            "Insufficient cluster resources to launch trial: "
+                            "trial requested {} but the cluster only has {} "
+                            "available.").format(
+                                trial.resources.summary_string(),
+                                self._avail_resources.summary_string()))
                 elif trial.status == Trial.PAUSED:
                     raise TuneError(
                         "There are paused trials, but no more pending "