def test_saving_agent():
print()
# Different agents
dqn_agent = TaskPricingDqnAgent(0,
create_lstm_dqn_network(9, 10),
save_folder='tmp')
ddpg_agent = TaskPricingDdpgAgent(1,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
save_folder='tmp')
td3_agent = TaskPricingTD3Agent(2,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
create_lstm_critic_network(9),
save_folder='tmp')
# Save the agent
dqn_agent.save('agent/checkpoints/')
ddpg_agent.save('agent/checkpoints/')
td3_agent.save('agent/checkpoints/')
# Check that loading works
loaded_model = create_lstm_dqn_network(9, 10)
loaded_model.load_weights(
f'agent/checkpoints/tmp/Task_pricing_Dqn_agent_0/update_0')
assert all(
tf.reduce_all(weights == load_weights) for weights, load_weights in
zip(dqn_agent.model_network.variables, loaded_model.variables))
def test_train_agents():
print()
setup_tensorboard('training/tmp/', 'train_agents')
env = OnlineFlexibleResourceAllocationEnv('training/settings/basic.env')
pricing_agents = [
TaskPricingDqnAgent(0,
create_rnn_dqn_network(9, 5),
batch_size=16,
initial_training_replay_size=16,
training_freq=100),
]
weighting_agents = [
ResourceWeightingDqnAgent(2,
create_rnn_dqn_network(16, 5),
batch_size=16,
initial_training_replay_size=16,
training_freq=100)
]
for _ in range(2):
train_agent(env, pricing_agents, weighting_agents)
# noinspection PyTypeChecker
agents: List[
ReinforcementLearningAgent] = pricing_agents + weighting_agents
for agent in agents:
assert 0 < agent.total_updates
def test_agent_actions():
print()
pricing_agents = [
TaskPricingDqnAgent(0, create_lstm_dqn_network(9, 5)),
TaskPricingDdqnAgent(1, create_lstm_dqn_network(9, 5)),
TaskPricingDuelingDqnAgent(2, create_lstm_dueling_dqn_network(9, 5)),
TaskPricingCategoricalDqnAgent(
3, create_lstm_categorical_dqn_network(9, 5)),
TaskPricingDdpgAgent(4, create_lstm_actor_network(9),
create_lstm_critic_network(9)),
TaskPricingTD3Agent(5, create_lstm_actor_network(9),
create_lstm_critic_network(9),
create_lstm_critic_network(9))
]
weighting_agents = [
ResourceWeightingDqnAgent(0, create_lstm_dqn_network(16, 5)),
ResourceWeightingDdqnAgent(1, create_lstm_dqn_network(16, 5)),
ResourceWeightingDuelingDqnAgent(
2, create_lstm_dueling_dqn_network(16, 5)),
ResourceWeightingCategoricalDqnAgent(
3, create_lstm_categorical_dqn_network(16, 5)),
ResourceWeightingDdpgAgent(4, create_lstm_actor_network(16),
create_lstm_critic_network(16)),
ResourceWeightingTD3Agent(5, create_lstm_actor_network(16),
create_lstm_critic_network(16),
create_lstm_critic_network(16))
]
env, state = OnlineFlexibleResourceAllocationEnv.load_env(
'agent/settings/actions.env')
for agent in pricing_agents:
actions = {
server: agent.bid(state.auction_task, tasks, server,
state.time_step)
for server, tasks in state.server_tasks.items()
}
# noinspection PyUnboundLocalVariable
print(
f'Actions: {{{", ".join([f"{server.name}: {action}" for server, action in actions.items()])}}}'
)
state, rewards, done, _ = env.step(actions)
for agent in weighting_agents:
actions = {
server: agent.weight(tasks, server, state.time_step)
for server, tasks in state.server_tasks.items()
}
print(
f'Actions: {{{", ".join([f"{server.name}: {list(task_action.values())}" for server, task_action in actions.items()])}}}'
)
state, rewards, done, _ = env.step(actions)
def test_agent_evaluation():
print()
setup_tensorboard('training/results/tmp/', 'agent_eval')
env = OnlineFlexibleResourceAllocationEnv('training/settings/basic.env')
eval_envs = generate_eval_envs(env,
5,
'training/settings/tmp/',
overwrite=True)
assert len(os.listdir('training/settings/tmp/')) == 5
total_auctions, total_resource_allocation = 0, 0
for eval_env in eval_envs:
env, state = OnlineFlexibleResourceAllocationEnv.load_env(eval_env)
total_auctions += len(env._unallocated_tasks) + (
1 if state.auction_task is not None else 0)
total_resource_allocation += env._total_time_steps + 1
pricing_agents = [
TaskPricingDqnAgent(0, create_bidirectional_dqn_network(9, 5)),
TaskPricingDdpgAgent(1, create_lstm_actor_network(9),
create_lstm_critic_network(9))
]
weighting_agents = [
ResourceWeightingDqnAgent(2, create_bidirectional_dqn_network(16, 5)),
ResourceWeightingDdpgAgent(3, create_lstm_actor_network(16),
create_lstm_critic_network(16)),
]
results = eval_agent(eval_envs, 0, pricing_agents, weighting_agents)
print(
f'Results - Total prices: {results.total_prices}, Number of completed tasks: {results.num_completed_tasks}, '
f'failed tasks: {results.num_failed_tasks}, winning prices: {results.winning_prices}, '
f'Number of auctions: {results.num_auctions}, resource allocations: {results.num_resource_allocations}'
)
assert 0 < results.num_completed_tasks
assert 0 < results.num_failed_tasks
assert results.num_auctions == total_auctions
assert results.num_resource_allocations == total_resource_allocation
def load_agents():
task_pricing_agents = [
TaskPricingDqnAgent(agent_num, create_lstm_dqn_network(9, 21))
for agent_num in range(3)
]
task_pricing_agents[0].model_network.load_weights(
'./analysis/fixed_heuristics/eval_agents/Task_pricing_Dqn_agent_0/update_80922'
)
task_pricing_agents[1].model_network.load_weights(
'./analysis/fixed_heuristics/eval_agents/Task_pricing_Dqn_agent_1/update_86909'
)
task_pricing_agents[2].model_network.load_weights(
'./analysis/fixed_heuristics/eval_agents/Task_pricing_Dqn_agent_2/update_88937'
)
resource_weighting_agents = [
ResourceWeightingDqnAgent(0, create_lstm_dqn_network(16, 11))
]
resource_weighting_agents[0].model_network.load_weights(
'./analysis/fixed_heuristics/eval_agents/Resource_weighting_Dqn_agent_0/update_440898'
)
return task_pricing_agents, resource_weighting_agents
writer, datetime = setup_tensorboard('training/results/logs/', folder)
save_folder = f'{folder}_{datetime}'
env = OnlineFlexibleResourceAllocationEnv([
'./training/settings/basic.env',
'./training/settings/large_tasks_servers.env',
'./training/settings/limited_resources.env',
'./training/settings/mixture_tasks_servers.env'
])
eval_envs = generate_eval_envs(
env, 20, f'./training/settings/eval_envs/network_arch/')
task_pricing_agents = [
TaskPricingDqnAgent(agent_num,
create_bidirectional_dqn_network(9, 21),
save_folder=save_folder) for agent_num in range(3)
]
resource_weighting_agents = [
ResourceWeightingDqnAgent(0,
create_bidirectional_dqn_network(16, 11),
save_folder=save_folder)
]
# Train the agents
with writer.as_default():
run_training(env, eval_envs, 600, task_pricing_agents,
resource_weighting_agents, 10)
for agent in task_pricing_agents:
agent.save()
writer, datetime = setup_tensorboard('training/results/logs/', folder)
save_folder = f'{folder}_{datetime}'
env = OnlineFlexibleResourceAllocationEnv([
'./training/settings/basic.env',
'./training/settings/large_tasks_servers.env',
'./training/settings/limited_resources.env',
'./training/settings/mixture_tasks_servers.env'
])
eval_envs = generate_eval_envs(
env, 20, f'./training/settings/eval_envs/network_arch/')
task_pricing_agents = [
TaskPricingDqnAgent(agent_num,
create_lstm_dqn_network(9, 21),
save_folder=save_folder) for agent_num in range(3)
]
resource_weighting_agents = [
ResourceWeightingSeq2SeqAgent(0,
create_seq2seq_actor_network(),
create_seq2seq_critic_network(),
create_seq2seq_critic_network(),
save_folder=save_folder)
]
with writer.as_default():
run_training(env, eval_envs, 600, task_pricing_agents,
resource_weighting_agents, 10)
for agent in task_pricing_agents:
def test_networks():
print()
# Environment setup
auction_task = Task('Test 4', 69.0, 35.0, 10.0, 0, 12)
tasks = [
Task('Test 1', 76.0, 36.0, 16.0, 0, 12, TaskStage.LOADING, 50.0, price=1),
Task('Test 2', 75.0, 37.0, 12.0, 0, 12, TaskStage.COMPUTING, 75.0, 10.0, price=1),
Task('Test 3', 72.0, 47.0, 20.0, 0, 7, TaskStage.COMPUTING, 72.0, 25.0, price=1)
]
server = Server('Test', 220.0, 35.0, 22.0)
# Assert the environment is valid
auction_task.assert_valid()
for task in tasks:
task.assert_valid()
server.assert_valid()
# List of networks
pricing_networks = [
create_bidirectional_dqn_network(9, 3),
create_lstm_dqn_network(9, 3),
create_gru_dqn_network(9, 3),
create_rnn_dqn_network(9, 3),
create_lstm_dueling_dqn_network(9, 3),
create_lstm_dueling_dqn_network(9, 3, combiner='max')
]
weighting_networks = [
create_bidirectional_dqn_network(16, 3),
create_lstm_dqn_network(16, 3),
create_gru_dqn_network(16, 3),
create_rnn_dqn_network(16, 3),
create_lstm_dueling_dqn_network(16, 3),
create_lstm_dueling_dqn_network(16, 3, combiner='max')
]
# Network observations
auction_obs = tf.expand_dims(TaskPricingDqnAgent._network_obs(auction_task, tasks, server, 0), axis=0)
resource_obs = tf.expand_dims(ResourceWeightingDqnAgent._network_obs(tasks[0], tasks, server, 0), axis=0)
print(f'Auction obs: {auction_obs}')
print(f'Resource allocation obs: {resource_obs}')
# Loop over the networks to find the output and output shape is correct
for pricing_network, weighting_network in zip(pricing_networks, weighting_networks):
auction_output = pricing_network(auction_obs)
resource_output = weighting_network(resource_obs)
print(f'Network: {pricing_network.name}'
f'\n\tAuction: {auction_output} ({auction_output.shape})'
f'\n\tResource allocation: {resource_output} ({resource_output.shape})')
assert auction_output.shape == (1, 3)
assert resource_output.shape == (1, 3)
# Check for the categorical dqn networks as it is a special case
pricing_network = create_lstm_categorical_dqn_network(9, 3, num_atoms=10)
weighting_network = create_lstm_categorical_dqn_network(16, 3, num_atoms=10)
auction_output = pricing_network(auction_obs)
resource_output = weighting_network(resource_obs)
print(f'Network: {pricing_network.name}'
f'\n\tAuction: {auction_output}'
f'\n\tResource allocation: {resource_output}')
assert auction_output.shape == (1, 3, 10)
assert resource_output.shape == (1, 3, 10)
# Check for the ddpg networks as it is a special case
pricing_actor_networks = [
create_lstm_actor_network(9)
]
pricing_critic_networks = [
create_lstm_critic_network(9)
]
weighting_actor_networks = [
create_lstm_actor_network(16)
]
weighting_critic_networks = [
create_lstm_critic_network(16)
]
for pricing_actor, pricing_critic, weighting_actor, weighting_critic in zip(
pricing_actor_networks, pricing_critic_networks, weighting_actor_networks, weighting_critic_networks):
auction_actor_output = pricing_actor(auction_obs)
weighting_actor_output = weighting_actor(resource_obs)
auction_critic_output = pricing_critic([auction_obs, auction_actor_output])
weighting_critic_output = weighting_critic([resource_obs, weighting_actor_output])
print(f'Network - actor: {pricing_actor.name}, critic: {pricing_critic.name}'
f'\n\tAuction Actor: {auction_actor_output}, critic: {auction_critic_output}'
f'\n\tWeighting Actor: {weighting_actor_output}, critic: {weighting_critic_output}')
assert auction_actor_output.shape == (1, 1)
assert weighting_actor_output.shape == (1, 1)
assert auction_critic_output.shape == (1, 1)
assert weighting_critic_output.shape == (1, 1)
def test_task_price_training():
print()
setup_tensorboard('/tmp/results/', 'price_training')
# List of agents
agents: List[TaskPricingRLAgent] = [
TaskPricingDqnAgent(0,
create_lstm_dqn_network(9, 10),
batch_size=4,
save_folder='tmp'),
TaskPricingDdqnAgent(1,
create_lstm_dqn_network(9, 10),
batch_size=4,
save_folder='tmp'),
TaskPricingDuelingDqnAgent(2,
create_lstm_dueling_dqn_network(9, 10),
batch_size=4,
save_folder='tmp'),
TaskPricingCategoricalDqnAgent(3,
create_lstm_categorical_dqn_network(
9, 10),
batch_size=4,
save_folder='tmp'),
TaskPricingDdpgAgent(4,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
batch_size=4,
save_folder='tmp'),
TaskPricingTD3Agent(5,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
create_lstm_critic_network(9),
batch_size=4,
save_folder='tmp')
]
# Load the environment
env, state = OnlineFlexibleResourceAllocationEnv.load_env(
'training/settings/auction.env')
# Servers
server_1, server_2 = list(state.server_tasks.keys())
# Actions
actions = {server_1: 1.0, server_2: 2.0}
# Environment step
next_state, reward, done, info = env.step(actions)
# Server states
server_1_state = TaskPricingState(state.auction_task,
state.server_tasks[server_1], server_1,
state.time_step)
server_2_state = TaskPricingState(state.auction_task,
state.server_tasks[server_2], server_2,
state.time_step)
# Next server states
next_server_1_state = TaskPricingState(next_state.auction_task,
next_state.server_tasks[server_1],
server_1, next_state.time_step)
next_server_2_state = TaskPricingState(next_state.auction_task,
next_state.server_tasks[server_2],
server_2, next_state.time_step)
# Finished auction task
finished_task = next(finished_task
for finished_task in next_state.server_tasks[server_1]
if finished_task == state.auction_task)
finished_task = finished_task._replace(stage=TaskStage.COMPLETED)
failed_task = finished_task._replace(stage=TaskStage.FAILED)
# Loop over the agents, add the observations and try training
for agent in agents:
agent.winning_auction_bid(server_1_state, actions[server_1],
finished_task, next_server_1_state)
agent.winning_auction_bid(server_1_state, actions[server_1],
failed_task, next_server_1_state)
agent.failed_auction_bid(server_2_state, actions[server_2],
next_server_2_state)
agent.failed_auction_bid(server_2_state, 0, next_server_2_state)
agent.train()
print(
f'Rewards: {[trajectory[3] for trajectory in agents[0].replay_buffer]}'
)
Training of single agents with multiple environments
"""
from __future__ import annotations
from agents.rl_agents.agents.dqn import ResourceWeightingDqnAgent, TaskPricingDqnAgent
from agents.rl_agents.neural_networks.dqn_networks import create_lstm_dqn_network
from training.train_agents import setup_tensorboard, multi_env_single_env_training
if __name__ == "__main__":
folder = 'single_agent_multi_envs'
primary_writer, datetime = setup_tensorboard('training/results/logs/',
folder)
save_folder = f'{folder}_{datetime}'
task_pricing_agents = [
TaskPricingDqnAgent(0,
create_lstm_dqn_network(9, 21),
save_folder=save_folder)
]
resource_weighting_agents = [
ResourceWeightingDqnAgent(0,
create_lstm_dqn_network(16, 11),
save_folder=save_folder)
]
multi_env_single_env_training(folder, datetime, primary_writer,
task_pricing_agents,
resource_weighting_agents)
def test_epsilon_policy():
print()
# Tests the epsilon policy by getting agent actions that should update the agent epsilon over time
env, state = OnlineFlexibleResourceAllocationEnv.load_env(
'agent/settings/actions.env')
# Number of epsilon steps for the agents
epsilon_steps = 25
# Agents that have a custom _get_action function
pricing_agents = [
TaskPricingDqnAgent(0,
create_lstm_dqn_network(9, 5),
epsilon_steps=epsilon_steps,
epsilon_update_freq=1,
epsilon_log_freq=1),
TaskPricingCategoricalDqnAgent(1,
create_lstm_categorical_dqn_network(
9, 5),
epsilon_steps=epsilon_steps,
epsilon_update_freq=1,
epsilon_log_freq=1),
TaskPricingDdpgAgent(2,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
epsilon_steps=epsilon_steps,
epsilon_update_freq=1,
epsilon_log_freq=1)
]
weighting_agents = [
ResourceWeightingDqnAgent(0,
create_lstm_dqn_network(16, 5),
epsilon_steps=epsilon_steps,
epsilon_update_freq=1,
epsilon_log_freq=1),
ResourceWeightingCategoricalDqnAgent(
1,
create_lstm_categorical_dqn_network(16, 5),
epsilon_steps=epsilon_steps,
epsilon_update_freq=1,
epsilon_log_freq=1),
ResourceWeightingDdpgAgent(2,
create_lstm_actor_network(16),
create_lstm_critic_network(16),
epsilon_steps=epsilon_steps,
epsilon_update_freq=1,
epsilon_log_freq=1)
]
# Generate a tf writer and generate actions that will update the epsilon values for both agents
writer = tf.summary.create_file_writer(f'agent/tmp/testing_epsilon')
num_steps = 10
with writer.as_default():
for _ in range(num_steps):
for agent in pricing_agents:
actions = {
server: agent.bid(state.auction_task,
tasks,
server,
state.time_step,
training=True)
for server, tasks in state.server_tasks.items()
}
state, rewards, done, _ = env.step(actions)
for _ in range(num_steps):
for agent in weighting_agents:
actions = {
server: agent.weight(tasks,
server,
state.time_step,
training=True)
for server, tasks in state.server_tasks.items()
}
state, rewards, done, _ = env.step(actions)
# Check that the resulting total action are valid
for agent in pricing_agents:
print(f'Agent: {agent.name}')
assert agent.total_actions == num_steps * 3
for agent in weighting_agents:
print(f'Agent: {agent.name}')
assert agent.total_actions == num_steps * 3
# Check that the agent epsilon are correct
assert pricing_agents[0].final_epsilon == pricing_agents[
0].epsilon and pricing_agents[1].final_epsilon == pricing_agents[
1].epsilon
assert weighting_agents[0].final_epsilon == weighting_agents[
0].epsilon and weighting_agents[1].final_epsilon == weighting_agents[
1].epsilon
assert pricing_agents[2].final_epsilon_std == pricing_agents[2].epsilon_std
assert weighting_agents[2].final_epsilon_std == weighting_agents[
2].epsilon_std
def test_build_agent():
def assert_args(test_agent, args):
"""
Asserts that the proposed arguments have assigned to the agent
Args:
test_agent: The test agent
args: The argument used on the agent
"""
for arg_name, arg_value in args.items():
assert getattr(test_agent, arg_name) == arg_value, \
f'Attr: {arg_name}, correct value: {arg_value}, actual value: {getattr(test_agent, arg_name)}'
# Check inheritance arguments
reinforcement_learning_arguments = {
'batch_size': 16,
'error_loss_fn': tf.compat.v1.losses.mean_squared_error,
'initial_training_replay_size': 1000,
'training_freq': 2,
'replay_buffer_length': 20000,
'save_frequency': 12500,
'save_folder': 'test',
'discount_factor': 0.9
}
dqn_arguments = {
'target_update_tau': 1.0,
'target_update_frequency': 2500,
'optimiser': tf.keras.optimizers.Adadelta(),
'initial_epsilon': 0.5,
'final_epsilon': 0.2,
'epsilon_update_freq': 25,
'epsilon_log_freq': 10,
}
ddpg_arguments = {
'actor_optimiser': tf.keras.optimizers.Adadelta(),
'critic_optimiser': tf.keras.optimizers.Adadelta(),
'initial_epsilon_std': 0.8,
'final_epsilon_std': 0.1,
'epsilon_update_freq': 25,
'epsilon_log_freq': 10,
'min_value': -15.0,
'max_value': 15
}
pricing_arguments = {
'failed_auction_reward': -100,
'failed_multiplier': -100
}
weighting_arguments = {
'other_task_discount': 0.2,
'success_reward': 1,
'failed_reward': -2
}
# DQN Agent arguments ----------------------------------------------------------------------
dqn_pricing_arguments = {
**reinforcement_learning_arguments,
**dqn_arguments,
**pricing_arguments
}
dqn_weighting_arguments = {
**reinforcement_learning_arguments,
**dqn_arguments,
**weighting_arguments
}
pricing_network = create_lstm_dqn_network(9, 10)
categorical_pricing_network = create_lstm_categorical_dqn_network(9, 10)
pricing_agents = [
TaskPricingDqnAgent(0, pricing_network, **dqn_pricing_arguments),
TaskPricingDdqnAgent(1, pricing_network, **dqn_pricing_arguments),
TaskPricingDuelingDqnAgent(2, pricing_network,
**dqn_pricing_arguments),
TaskPricingCategoricalDqnAgent(3, categorical_pricing_network,
**dqn_pricing_arguments)
]
for agent in pricing_agents:
print(f'Agent: {agent.name}')
assert_args(agent, dqn_pricing_arguments)
weighting_network = create_lstm_dqn_network(16, 10)
categorical_weighting_network = create_lstm_categorical_dqn_network(16, 10)
weighting_agents = [
ResourceWeightingDqnAgent(0, weighting_network,
**dqn_weighting_arguments),
ResourceWeightingDdqnAgent(1, weighting_network,
**dqn_weighting_arguments),
ResourceWeightingDuelingDqnAgent(2, weighting_network,
**dqn_weighting_arguments),
ResourceWeightingCategoricalDqnAgent(3, categorical_weighting_network,
**dqn_weighting_arguments)
]
for agent in weighting_agents:
print(f'Agent: {agent.name}')
assert_args(agent, dqn_weighting_arguments)
# PG Agent arguments ----------------------------------------------------------------------------------
ddpg_pricing_arguments = {
**reinforcement_learning_arguments,
**ddpg_arguments,
**pricing_arguments
}
ddpg_weighting_arguments = {
**reinforcement_learning_arguments,
**ddpg_arguments,
**weighting_arguments
}
pricing_agents = [
TaskPricingDdpgAgent(3, create_lstm_actor_network(9),
create_lstm_critic_network(9),
**ddpg_pricing_arguments),
TaskPricingTD3Agent(4, create_lstm_actor_network(9),
create_lstm_critic_network(9),
create_lstm_critic_network(9),
**ddpg_pricing_arguments)
]
for agent in pricing_agents:
print(f'Agent: {agent.name}')
assert_args(agent, ddpg_pricing_arguments)
weighting_agents = [
ResourceWeightingDdpgAgent(3, create_lstm_actor_network(16),
create_lstm_critic_network(16),
**ddpg_weighting_arguments),
ResourceWeightingTD3Agent(4, create_lstm_actor_network(16),
create_lstm_critic_network(16),
create_lstm_critic_network(16),
**ddpg_weighting_arguments)
]
for agent in weighting_agents:
print(f'Agent: {agent.name}')
assert_args(agent, ddpg_weighting_arguments)