from ddpg import Agent
from OUNoise import Noise
import matplotlib.pyplot as plt
env = PidEnv(setpoint=20)
batch_size = 128
rewards = []
agent = Agent(num_states=5, num_actions=3)
noise = Noise(num_actions=3)
for episode in range(30):
state = env.reset()
noise.reset()
eps_reward = 0
for step in range(500):
action = agent.get_action(state)
action = noise.get_action(action, step)
new_state, reward = env.step(action)
agent.mem.push(state, action, reward, new_state)
agent.learn(batch_size)
state = new_state
eps_reward += reward
rewards.append(eps_reward)
plt.plot(rewards)
setpoints.append(curr)
agent.metalearn(setpoints)
for episode in range(1):
noise.reset()
eps_reward = 0
print(episode)
best_parameters = [(0,0,0), 0, 0]
step = 0
for i in range(total_steps):
print(step)
step += 1
setpoint = 20 if random == False else np.random.random()*100
state = env.reset(setpoint)
action_before = agent.get_action(state)
action = noise.get_action(action_before, step)
if np.isnan(action[0]).any():
print(action)
print(action_before)
print(state)
for i in agent.actor.parameters():
print(i)
exit = True
break
new_state, reward = env.step(action)
if reward > -1000000:
agent.mem.push(state, action, reward, new_state)
if reward > -100:
normalized.append(reward)
from utils import fetch_protein
from protein import ProteinState
from ddpg import Agent, ReplayBuffer
EPISODES = 10000
STEPS = 500
if __name__ == "__main__":
goal_state = fetch_protein("2jof")
state_dim = goal_state.n_residues() * 2
action_dim = goal_state.n_residues() * 2
buffer = ReplayBuffer(10000)
agent = Agent(state_dim, action_dim, (0, 360))
for _ in range(EPISODES):
data = {"state": ProteinState(n_residues=goal_state.n_residues())}
for _ in range(STEPS):
action = agent.get_action(data["state"])
next_state = data["state"].do_action(action)
reward = data["state"].eval_state() - next_state.eval_state()
buffer.append(data["state"], action, reward, next_state)
agent.update(buffer)
print(data["state"].l2_norm(goal_state))
data["state"] = ProteinState(angles=next_state.angles())