From cabf5514156ea888ac627e8b0ea04682521999c8 Mon Sep 17 00:00:00 2001
From: "Egli Adrian (IT-SCI-API-PFI)" <adrian.egli@sbb.ch>
Date: Fri, 6 Nov 2020 14:19:09 +0100
Subject: [PATCH] DQN & PPO
---
reinforcement_learning/ppo/ppo_agent.py | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/reinforcement_learning/ppo/ppo_agent.py b/reinforcement_learning/ppo/ppo_agent.py
index a7431f8..350119a 100644
--- a/reinforcement_learning/ppo/ppo_agent.py
+++ b/reinforcement_learning/ppo/ppo_agent.py
@@ -39,6 +39,11 @@ class PPOAgent(Policy):
# Decide on an action to take in the environment
def act(self, state, eps=None):
+ if eps is not None:
+ # Epsilon-greedy action selection
+ if np.random.random() < eps:
+ return np.random.choice(np.arange(self.action_size))
+
self.policy.eval()
with torch.no_grad():
output = self.policy(torch.from_numpy(state).float().unsqueeze(0).to(device))
--
GitLab