#!/usr/bin/env python3
import gym
import sys
sys.path.insert(0, "../../../spyrl")
from spyrl.activity.learning import Learning
from spyrl.activity.activity_config import ActivityConfig
from spyrl.agent_builder.impl.actor_critic_traces_agent_builder import ActorCriticTracesAgentBuilder
from spyrl.listener.impl.basic_functions import BasicFunctions
from spyrl.listener.impl.file_log_listener import RewardType
from example.lunarlander.helper.lunarlander_discretiser import LunarLanderDiscretiser24576
if __name__ == '__main__':
env = gym.make('LunarLander-v2')
num_actions = env.action_space.n
config = ActivityConfig(start_trial=1,
num_trials=1,
num_episodes=100,
out_path='results/acet-21/')
agent_builder = ActorCriticTracesAgentBuilder(
num_actions, discretiser=LunarLanderDiscretiser24576())
milestone_episodes = [1000, 2000, 4000, 5000, 8000]
learning = Learning(
listener=BasicFunctions(render=False,
draw=False,
reward_type=RewardType.TOTAL,
milestone_episodes=milestone_episodes))
learning.learn(env, agent_builder, config)
def create_agent(self, seed, initial_policy_path=None):
num_inputs = 8
memory_size = 50_000
batch_size = 64
dqn_dims = [num_inputs, 128, self.num_actions]
c = 100
return DQNWithTargetNetworkAgent(memory_size, batch_size, dqn_dims,
self.normaliser, c, seed)
if __name__ == '__main__':
start_trial = 1
if len(sys.argv) > 1:
start_trial = int(sys.argv[1])
env = gym.make('LunarLander-v2')
config = ActivityConfig(start_trial=start_trial,
num_trials=10,
num_episodes=10000,
out_path='result/lunarlander/dqnwtn-03/')
agent_builder = DQNAgentBuilder(env.action_space.n)
milestone_episodes = [1000, 2000, 3000, 4000, 5000, 6000, 7000, 8000, 9000]
learning = Learning(
listener=BasicFunctions(render=False,
draw=False,
milestone_episodes=milestone_episodes,
reward_type=RewardType.TOTAL))
learning.add_listener(Gmailer("DQNwTN-03"))
learning.learn(env, agent_builder, config)
class DoubleD2DSQLAgentBuilder(AgentBuilder):
def create_agent(self, seed, initial_policy_path=None):
num_inputs = 8
memory_size = 50_000; batch_size = 64; dqn_dims = [num_inputs, 128, self.num_actions]
trial = seed
normalised_training_set_path = os.path.join(get_project_dir(), normalised_training_set_parent_path + \
'/d2dspl-normalised_training_set-' + str(trial).zfill(2) + '-00005000.txt')
return DoubleD2DSQLAgent2(normalised_training_set_path, target_loss, memory_size, batch_size, dqn_dims, self.normaliser, seed)
if __name__ == '__main__':
env = gym.make('LunarLander-v2')
num_actions = env.action_space.n
max_num_samples_for_classifier = 500
num_episodes = 10000
session_id = '22'
target_loss = 0.01 #0.001
normalised_training_set_parent_path = 'result/lunarlander/d2dspl-acet-' + str(num_episodes) + '-' + session_id
description = 'dd2dsql-' + session_id + '. mem_size=50,000, batch_size=64, hidden dims=128, DoubleD2DSQLAgent2 (fixed epsilon)' + \
'training set from ' + normalised_training_set_parent_path + \
'/d2dspl-normalised_training_set-0x-00050000.txt\n' + 'target_loss: ' + str(target_loss)
milestone_episodes = [1000,2000,3000,4000,5000,6000,7000,8000,9000,10000]
config = ActivityConfig(start_trial=1, num_trials = 10, num_episodes=num_episodes,
out_path='result/lunarlander/dd2dsql-' + session_id + '/')
agent_builder = DoubleD2DSQLAgentBuilder(num_actions)
listeners = [BasicFunctions(render=False, draw=False, reward_type=RewardType.TOTAL, milestone_episodes=milestone_episodes),
Gmailer("DD2DSQL-22"), SessionLogger(description)]
learning = Learning(listeners=listeners)
learning.learn(env, agent_builder, config)
from spyrl.listener.impl.basic_functions import BasicFunctions
__author__ = "Budi Kurniawan"
__copyright__ = "Copyright 2021, Budi Kurniawan"
__license__ = "GPL"
__version__ = "0.1.0"
class PPOAgentBuilder(AgentBuilder):
def create_agent(self, seed, initial_policy_path=None):
normaliser = None
seed = 1
return PPOAgent(nn_dims, normaliser, seed, local_steps_per_epoch=7)
if __name__ == '__main__':
id = 'CartPole-v2'
gym.envs.register(id=id,
entry_point='gym.envs.classic_control:CartPoleEnv',
max_episode_steps=100_000)
env = gym.make(id)
num_actions = env.action_space.n
num_states = env.observation_space.shape[0]
nn_dims = (num_states, 64, 64, num_actions)
out_path = os.path.join(get_project_dir(), 'result/cartpole/ppo-test/')
config = ActivityConfig(num_episodes=2, out_path=out_path)
agent_builder = PPOAgentBuilder(num_actions)
learning = Learning(listener=BasicFunctions(render=False))
learning.learn(env, agent_builder, config)
from spyrl.listener.impl.file_log_listener import RewardType
from example.lunarlander.helper.lunarlander_discretiser import LunarLanderDiscretiser,\
LunarLanderDiscretiser24576
__author__ = "Budi Kurniawan"
__copyright__ = "Copyright 2021, Budi Kurniawan"
__license__ = "GPL"
__version__ = "0.1.0"
if __name__ == '__main__':
env = gym.make('LunarLander-v2')
num_actions = env.action_space.n
max_num_samples_for_classifier = 500
num_episodes = 10000
session_id = '24'
milestone_episodes = [5000, 10000]
config = ActivityConfig(start_trial=1,
num_trials=10,
num_episodes=num_episodes,
out_path='result/lunarlander/d2dspl-acet-' +
str(num_episodes) + '-' + session_id + '/')
agent_builder = D2DSPLActorCriticTracesAgentBuilder(
num_actions, LunarLanderDiscretiser24576(),
max_num_samples_for_classifier, None, [80, 80])
learning = Learning(
listener=BasicFunctions(render=False,
draw=False,
reward_type=RewardType.TOTAL,
milestone_episodes=milestone_episodes))
learning.add_listener(Gmailer("D2DSPL-ACET-10000-24"))
learning.learn(env, agent_builder, config)