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 test_resource_allocation_training():
print()
setup_tensorboard('/tmp/results/', 'resource_allocation_training')
# List of agents
agents: List[ResourceWeightingRLAgent] = [
ResourceWeightingDqnAgent(0,
create_lstm_dqn_network(16, 10),
batch_size=4,
save_folder='tmp'),
ResourceWeightingDdqnAgent(1,
create_lstm_dqn_network(16, 10),
batch_size=4,
save_folder='tmp'),
ResourceWeightingDuelingDqnAgent(2,
create_lstm_dueling_dqn_network(
16, 10),
batch_size=4,
save_folder='tmp'),
ResourceWeightingCategoricalDqnAgent(
3,
create_lstm_categorical_dqn_network(16, 10),
batch_size=2,
save_folder='tmp'),
ResourceWeightingDdpgAgent(4,
create_lstm_actor_network(16),
create_lstm_critic_network(16),
batch_size=4,
save_folder='tmp'),
ResourceWeightingTD3Agent(5,
create_lstm_actor_network(16),
create_lstm_critic_network(16),
create_lstm_critic_network(16),
batch_size=4,
save_folder='tmp'),
]
# Load the environment
env, state = OnlineFlexibleResourceAllocationEnv.load_env(
'training/settings/resource_allocation.env')
# Servers and tasks
server = list(state.server_tasks.keys())[0]
task_1, task_2, task_3, task_4 = list(state.server_tasks[server])
# Actions
actions = {server: {task_1: 1.0, task_2: 3.0, task_3: 0.0, task_4: 5.0}}
# Environment step
next_state, rewards, done, _ = env.step(actions)
# Resource state
resource_state = ResourceAllocationState(state.server_tasks[server],
server, state.time_step)
# Next server and resource state
next_resource_state = ResourceAllocationState(
next_state.server_tasks[server], server, next_state.time_step)
for agent in agents:
agent.resource_allocation_obs(resource_state, actions[server],
next_resource_state, rewards[server])
agent.train()
agent = ResourceWeightingSeq2SeqAgent(6,
create_seq2seq_actor_network(),
create_seq2seq_critic_network(),
create_seq2seq_critic_network(),
batch_size=2,
save_folder='tmp')
agent.resource_allocation_obs(resource_state, actions[server],
next_resource_state, rewards[server])
agent.resource_allocation_obs(resource_state, actions[server],
next_resource_state, rewards[server])
agent.train()
print(
f'Rewards: {[trajectory[3] for trajectory in agents[0].replay_buffer]}'
)
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_ddpg_actions():
print()
# Check that DDPG actions are valid
env, state = OnlineFlexibleResourceAllocationEnv.load_env(
'agent/settings/actions.env')
repeat, max_repeat = 0, 10
auction_actions = {}
while repeat <= max_repeat:
pricing_agent = TaskPricingDdpgAgent(3,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
initial_epsilon=0.5)
auction_actions = {
server: pricing_agent.bid(state.auction_task, tasks, server,
state.time_step)
for server, tasks in state.server_tasks.items()
}
print(f'Greedy actions: {list(auction_actions.values())}')
if any(0 < action for server, action in auction_actions.items()):
auction_actions = {
server: pricing_agent.bid(state.auction_task,
tasks,
server,
state.time_step,
training=True)
for server, tasks in state.server_tasks.items()
}
print(
f'Epsilon Greedy actions: {list(auction_actions.values())}\n')
if any(0 < action for server, action in auction_actions.items()):
break
elif repeat == max_repeat:
raise Exception()
else:
repeat += 1
states, rewards, dones, _ = env.step(auction_actions)
repeat, max_repeat = 0, 10
while repeat <= max_repeat:
weighting_agent = ResourceWeightingDdpgAgent(
3,
create_lstm_actor_network(16),
create_lstm_critic_network(16),
initial_epsilon=0.5)
weighting_actions = {
server: weighting_agent.weight(tasks, server, state.time_step)
for server, tasks in state.server_tasks.items()
}
print(
f'Greedy actions: {[list(actions.values()) for actions in weighting_actions.values()]}'
)
if any(0 < action
for server, task_actions in weighting_actions.items()
for task, action in task_actions.items()):
weighting_actions = {
server: weighting_agent.weight(tasks,
server,
state.time_step,
training=True)
for server, tasks in state.server_tasks.items()
}
print(
f'Greedy actions: {[list(actions.values()) for actions in weighting_actions.values()]}'
)
if any(0 < action
for server, task_actions in weighting_actions.items()
for task, action in task_actions.items()):
break
elif repeat == max_repeat:
raise Exception()
else:
repeat += 1
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)
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, 5,
f'./training/settings/eval_envs/algo/')
task_pricing_agents = [
TaskPricingDdpgAgent(agent_num,
create_lstm_actor_network(9),
create_lstm_critic_network(9),
save_folder=save_folder) for agent_num in range(3)
]
resource_weighting_agents = [
ResourceWeightingDdpgAgent(0,
create_lstm_actor_network(16),
create_lstm_critic_network(16),
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:
agent.save()
for agent in resource_weighting_agents:
agent.save()