reduce experiment time but increase curriculum

tasks/rl_hopper.py

@@ -233,13 +233,13 @@ def prepare_data_tuples(states, actions, rewards, num_layers, skip_steps):
         observation, info = env.reset(seed=random_seed)
 
         num_layers = len(cortex_models)
-        num_courses = 20
+        num_courses = 60
         for course in range(num_courses):
             logging.info(f"Course {course}")
             total_steps = 0
-            num_trials = 20000
+            num_trials = 2000
             print_steps = max(1, num_trials // 100)
-            epsilon = 1 - 0.5 * (course + 1) / num_courses
+            epsilon = 1 - 0.75 * (course + 1) / num_courses
 
             next_best_targets = np.zeros((len(goals), len(goals[0][0])), dtype=np.float32)
             next_best_target_diffs = np.ones((len(goals), 1), dtype=np.float32) * 1e4
@@ -287,7 +287,7 @@ def prepare_data_tuples(states, actions, rewards, num_layers, skip_steps):
             for trainer in trainers:
                 trainer.prepare_batch(max_mini_batch_size=16, max_learning_sequence=32)
 
-            loop_train(trainers, 100000)
+            loop_train(trainers, 20000)
 
             for trainer in trainers:
                 trainer.clear_batch()