def score_genome(genome, episodes, pbar):
# make gym
env = gym.make("CartPole-v1")
# Create an agent that solves the environment:
config = AgentConfig(env=env,
n_episodes=episodes,
target_average=475,
update_every=genome['updateEvery'],
batch_size=genome['batchSize'],
gamma=genome['gamma'],
learning_rate=genome['learningRate'],
fc1_units=genome['fc1_units'],
fc2_units=genome['fc2_units'],
tau=genome['tau'],
memory_size=genome['memory_size'],
deepq_double_learning=True,
deepq_dueling_networks=True,
per_active=True)
agent = DQNAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent, verbose=1, pbar=pbar)
# Run the agent
score, checkpoint = runner.run_agent()
# Close
env.close()
# Return
return score.best_score, checkpoint
def score_genome(genome, episodes, pbar):
# Make the gym
env = gym.make("Acrobot-v1")
# Create an agent that solves the environment:
# "Acrobot-v1 is an unsolved environment, which means it does not have a specified reward threshold at which it's considered solved.""
config = AgentConfig(env=env,
device="cpu",
n_episodes=episodes,
target_average=-10,
max_t=200,
gamma=genome['gamma'],
learning_rate=genome['learningRate'],
fc1_units=genome['fc1_units'])
agent = REINFORCEAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent, verbose=1, pbar=pbar)
# Run the agent
score, checkpoint = runner.run_single_probability_trajectory()
# Return
return score.best_score, checkpoint
def random_agent(self):
env = gym.make(self.gym_name) # Make the gym
agent = self.get_agent(AgentConfig(env=env)) # Create the agent
runner = Runner(
agent) # Create a runner that runs the agent in the environment
runner.random() # Run the agent
env.close()
def enjoy_agent(self, agent_type):
env = gym.make(self.gym_name) # Make the gym
agent = self.get_agent(AgentConfig(env=env)) # Create the agent
runner = Runner(
agent) # Create a runner that runs the agent in the environment
if agent_type == "evolve":
self.enjoy_checkpoint(
runner, 'Checkpoints/{}_evo.ch'.format(self.gym_name.lower()))
elif agent_type == "train":
self.enjoy_checkpoint(
runner,
'Checkpoints/{}_train.ch'.format(self.gym_name.lower()))
env.close()
def score_genome(genome, episodes, pbar):
# make gym
env = gym.make("LunarLanderContinuous-v2")
# Create an agent that solves the environment:
config = AgentConfig(env=env,
seed=42,
target_average=200,
n_episodes=episodes,
batch_size=genome['batchSize'],
gamma=genome['gamma'],
actor_learning_rate=genome['actor_learningRate'],
critic_learning_rate=genome['critic_learningRate'],
weight_decay=genome['weight_decay'],
update_every=genome['updateEvery'],
ou_noise_sigma=0.01,
fc1_units=genome['fc1_units'],
fc2_units=genome['fc2_units'],
memory_size=genome['memory_size'],
per_active=False,
batch_norm=False)
agent = DDPGAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent, verbose=1, pbar=pbar)
# Run the agent
score, checkpoint = runner.run_agent()
# Close
env.close()
# Return
return score.best_score, checkpoint
import gym
# Internals
from Agents.PPO import PPOAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
from Agents.Utils.parallelEnv import parallelEnv
# Create an agent that solves the environment:
envs = parallelEnv("PongDeterministic-v4", n=8, seed=1234)
# Create the agent
config = AgentConfig(
parallelEnv=envs,
n_episodes=1500,
learning_rate=1e-4,
fc1_units=256,
ACTION_0=4,
ACTION_1=5,
max_t=500)
agent = PPOAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent)
# Run the agent
score = runner.run_parallel_trajectories()
# Visualize the score
score.visualize()
import gym
from Agents.Common import Runner
from Agents.DDPG import DDPGAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
# Make the gym
env = gym.make("LunarLanderContinuous-v2")
# Create the agent
agent = DDPGAgent(AgentConfig(env=env))
# Create a runner that runs the agent in the environment
runner = Runner(agent)
# Run the agent
runner.enjoy_checkpoint('Checkpoints/lunarlander_v2_ddpg_evo.ch')
env.close()
import gym
from Agents.Common import Runner
from Agents.REINFORCE import REINFORCEAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
# Make the gym
env = gym.make("Acrobot-v1")
# Create an agent that solves the environment:
# "Acrobot-v1 is an unsolved environment, which means it does not have a specified reward threshold at which it's considered solved"
config = AgentConfig(env=env,
n_episodes=300,
target_average=-100,
max_t=200,
gamma=0.92,
fc1_units=512,
learning_rate=0.002)
agent = REINFORCEAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent,
save_best_score='Checkpoints/acrobot_v1_reinforce_train.ch')
# Run the agent
score, checkpoint = runner.run_single_probability_trajectory()
# Visualize the score
score.visualize()
env.close()
import gym
from Agents.Common import Runner
from Agents.REINFORCE import REINFORCEAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
# Make the gym
env = gym.make("Acrobot-v1")
# Create the agent
agent = REINFORCEAgent(AgentConfig(env=env))
# Create a runner that runs the agent in the environment
runner = Runner(agent)
# Run the agent
runner.random()
env.close()
import gym
from Agents.Common import Runner
from Agents.DQN import DQNAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
# Make the gym
env = gym.make("CartPole-v1")
# Create the agent
agent = DQNAgent(AgentConfig(env=env))
# Create a runner that runs the agent in the environment
runner = Runner(agent)
# Run the agent
runner.enjoy_checkpoint('Checkpoints/cartpole_v1_dqn_evo.ch')
env.close()
import gym
from Agents.Common import Runner
from Agents.DQN import DQNAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
# Make the gym
env = gym.make("CartPole-v1")
# Create an agent that solves the environment:
# - CartPole-v1 defines "solving" as getting average reward of 475.0 over 100 consecutive trials."
config = AgentConfig(env=env,
n_episodes=1000,
target_average=475,
update_every=1,
batch_size=64,
gamma=0.95,
learning_rate=1e-4,
deepq_double_learning=True,
deepq_dueling_networks=True,
per_active=True)
agent = DQNAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent, save_best_score='Checkpoints/cartpole_v1_dqn_train.ch')
# Run the agent
score, checkpoint = runner.run_agent()
# Visualize the score
score.visualize()
from Agents.DDPG import DDPGAgent
from Agents.Common import AgentConfig
from Agents.Common import Runner
# Create environment
env = gym.make("LunarLanderContinuous-v2")
# Create the agent
config = AgentConfig(env=env,
seed=42,
target_average=200,
n_episodes=500,
batch_norm=False,
batch_size=64,
gamma=0.98,
actor_learning_rate=1e-4,
critic_learning_rate=1e-3,
weight_decay=0,
update_every=1,
ou_noise_sigma=0.01,
fc1_units=400,
fc2_units=300,
per_active=False)
agent = DDPGAgent(config)
# Create a runner that runs the agent in the environment
runner = Runner(agent,
save_best_score='Checkpoints/lunarlander_v2_ddpg_train.ch')
# Run the agent