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
from training.train_agents import generate_eval_envs, run_training, setup_tensorboard
if __name__ == "__main__":
# Setup tensorboard
folder = 'bidirectional_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,
from env.environment import OnlineFlexibleResourceAllocationEnv
from training.train_agents import generate_eval_envs, run_training, setup_tensorboard
if __name__ == "__main__":
folder = 'c51_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/algo/')
task_pricing_agents = [
TaskPricingCategoricalDqnAgent(agent_num,
create_lstm_categorical_dqn_network(
9, 21),
save_folder=save_folder)
for agent_num in range(3)
]
resource_weighting_agents = [
ResourceWeightingCategoricalDqnAgent(
0,
create_lstm_categorical_dqn_network(16, 11),
save_folder=save_folder)
]
if __name__ == "__main__":
eval_env = OnlineFlexibleResourceAllocationEnv([
'./analysis/fixed_heuristics/settings/basic.env',
'./analysis/fixed_heuristics/settings/large_tasks_servers.env',
'./analysis/fixed_heuristics/settings/limited_resources.env',
'./analysis/fixed_heuristics/settings/mixture_tasks_servers.env'
])
eval_pos = 0
while True:
if not os.path.exists(
f'./analysis/fixed_heuristics/eval_envs_{eval_pos}/'):
eval_envs = generate_eval_envs(
eval_env, 20,
f'./analysis/fixed_heuristics/eval_envs_{eval_pos}/')
break
else:
eval_pos += 1
task_pricing_agents, resource_weighting_agents = load_agents()
agent_results = eval_agent(eval_envs, 0, task_pricing_agents,
resource_weighting_agents)
print('Agent results')
print(f'Env completed tasks: {agent_results.env_completed_tasks}')
print(f'Env failed tasks: {agent_results.env_failed_tasks}')
print(f'Env attempted tasks: {agent_results.env_attempted_tasks}\n')
fixed_results = eval_fixed_env(eval_envs)
print(f'Fixed results env completed tasks: {fixed_results}')