Merge branch 'master' into runpod-rl_hopper

PVirie · Dec 25, 2024 · 198cda2 · 198cda2
2 parents b0be8c1 + 0930249
commit 198cda2
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/tasks/rl_hopper.py b/tasks/rl_hopper.py
@@ -239,7 +239,7 @@ def prepare_data_tuples(states, actions, rewards, num_layers, skip_steps):
             total_steps = 0
             num_trials = 2000
             print_steps = max(1, num_trials // 100)
-            epsilon = 0.2 + 0.7 * (course + 1) / num_courses
+            epsilon = 1 - 0.5 * (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
@@ -258,7 +258,7 @@ def prepare_data_tuples(states, actions, rewards, num_layers, skip_steps):
                 actions = []
                 rewards = []
                 for _ in range(200):
-                    if random.random() >= epsilon or course == 0:
+                    if random.random() <= epsilon or course == 0:
                         selected_action = env.action_space.sample()
                     else:
                         a = model.react(alg.State(observation.data), stable_state)