diff --git a/config.yaml b/config.yaml
index 2bf65d85..4512b9e7 100755
--- a/config.yaml
+++ b/config.yaml
@@ -20,7 +20,7 @@ train_args:
     maximum_episodes: 100000
     epochs: -1
     num_batchers: 2
-    eval_rate: 0.1
+    eval_coef: 0.85
     worker:
         num_parallel: 6
     lambda: 0.7
diff --git a/handyrl/train.py b/handyrl/train.py
index 75c0a31b..0e96f4fc 100755
--- a/handyrl/train.py
+++ b/handyrl/train.py
@@ -403,8 +403,8 @@ def __init__(self, args, net=None, remote=False):
         random.seed(args['seed'])
 
         self.env = make_env(env_args)
-        eval_modify_rate = (args['update_episodes'] ** 0.85) / args['update_episodes']
-        self.eval_rate = max(args['eval_rate'], eval_modify_rate)
+        self.eval_rate_fn = lambda n: (n ** self.args['eval_coef']) / n
+        self.eval_rate = self.eval_rate_fn(self.args['minimum_episodes'] + self.args['update_episodes'])
         self.shutdown_flag = False
         self.flags = set()
 
@@ -520,11 +520,17 @@ def output_wp(name, results):
             std = (r2 / (n + 1e-6) - mean ** 2) ** 0.5
             print('generation stats = %.3f +- %.3f' % (mean, std))
 
+        if self.model_epoch == 0:
+            self.num_episodes = 0
+            self.num_results = 0
+
         model, steps = self.trainer.update()
         if model is None:
             model = self.model
         self.update_model(model, steps)
 
+        # update evaluation ratio
+        self.eval_rate = self.eval_rate_fn(self.args['update_episodes'])
         # clear flags
         self.flags = set()