File size: 3,483 Bytes
a36c04c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 |
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from env.ev_charge_env import EVChargeEnv
class ActorCritic(nn.Module):
def __init__(self, obs_dim: int, act_dim: int):
super().__init__()
self.shared = nn.Sequential(
nn.Linear(obs_dim, 64),
nn.ReLU(),
nn.Linear(64, 64),
nn.ReLU(),
)
self.policy_head = nn.Linear(64, act_dim)
self.value_head = nn.Linear(64, 1)
def forward(self, x):
h = self.shared(x)
logits = self.policy_head(h)
value = self.value_head(h).squeeze(-1)
return logits, value
def make_env():
# You can change scenario here: "easy", "medium", "hard"
return EVChargeEnv(scenario="medium")
def run_episode(env, model, device, gamma=0.99):
obs, _ = env.reset()
obs = torch.tensor(obs, dtype=torch.float32, device=device)
log_probs = []
values = []
rewards = []
done = False
while not done:
logits, value = model(obs.unsqueeze(0)) # [1, obs_dim]
# Gaussian policy for continuous action in [0, 1]
mean = torch.sigmoid(logits.squeeze(0)) # [act_dim]
std = torch.ones_like(mean) * 0.2 # fixed std
dist = torch.distributions.Normal(mean, std)
action = dist.sample()
action_clipped = torch.clamp(action, 0.0, 1.0)
log_prob = dist.log_prob(action).sum()
np_action = action_clipped.detach().cpu().numpy()
next_obs, reward, terminated, truncated, _ = env.step(np_action)
log_probs.append(log_prob)
values.append(value)
rewards.append(torch.tensor(reward, dtype=torch.float32, device=device))
done = terminated or truncated
obs = torch.tensor(next_obs, dtype=torch.float32, device=device)
# Compute returns
returns = []
G = torch.tensor(0.0, device=device)
for r in reversed(rewards):
G = r + gamma * G
returns.insert(0, G)
returns = torch.stack(returns)
values = torch.stack(values).squeeze(-1)
log_probs = torch.stack(log_probs)
advantages = returns - values.detach()
policy_loss = -(log_probs * advantages).mean()
value_loss = (returns - values).pow(2).mean()
total_reward = float(sum(r.item() for r in rewards))
return policy_loss, value_loss, total_reward, len(rewards)
def train(num_episodes=200):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
env = make_env()
obs_dim = env.observation_space.shape[0]
act_dim = env.action_space.shape[0]
model = ActorCritic(obs_dim, act_dim).to(device)
optimizer = optim.Adam(model.parameters(), lr=3e-4)
reward_history = []
for episode in range(1, num_episodes + 1):
policy_loss, value_loss, total_reward, steps = run_episode(env, model, device)
loss = policy_loss + 0.5 * value_loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
reward_history.append(total_reward)
if episode % 10 == 0:
avg_last = np.mean(reward_history[-10:])
print(
f"Episode {episode:4d} | "
f"ep_reward={total_reward:.2f} | "
f"avg_last10={avg_last:.2f} | steps={steps}"
)
print("Training finished.")
print(f"Average reward over last 20 episodes: {np.mean(reward_history[-20:]):.2f}")
if __name__ == "__main__":
train(num_episodes=200)
|