def exp_run_industrial_datasets(iteration,
datasets=['paintcontrol', 'iofrol', 'gsdtsr']):
ags = [
lambda:
(agents.TableauAgent(histlen=retecs.DEFAULT_HISTORY_LENGTH,
learning_rate=retecs.DEFAULT_LEARNING_RATE,
state_size=retecs.DEFAULT_STATE_SIZE,
action_size=retecs.DEFAULT_NO_ACTIONS,
epsilon=retecs.DEFAULT_EPSILON), retecs.
preprocess_discrete, reward.timerank), lambda:
(agents.NetworkAgent(histlen=retecs.DEFAULT_HISTORY_LENGTH,
state_size=retecs.DEFAULT_STATE_SIZE,
action_size=1,
hidden_size=retecs.DEFAULT_NO_HIDDEN_NODES),
retecs.preprocess_continuous, reward.tcfail)
]
reward_funs = {
'failcount': reward.failcount,
'timerank': reward.timerank,
'tcfail': reward.tcfail
}
avg_napfd = []
for i, get_agent in enumerate(ags):
for sc in datasets:
for (reward_name, reward_fun) in list(reward_funs.items()):
agent, preprocessor, _ = get_agent()
file_appendix = 'rq_%s_%s_%s_%d' % (agent.name, sc,
reward_name, iteration)
scenario = get_scenario(sc)
rl_learning = retecs.PrioLearning(
agent=agent,
scenario_provider=scenario,
reward_function=reward_fun,
preprocess_function=preprocessor,
file_prefix=file_appendix,
dump_interval=100,
validation_interval=0,
output_dir=DATA_DIR)
res = rl_learning.train(no_scenarios=CI_CYCLES,
print_log=True,
plot_graphs=False,
save_graphs=False,
collect_comparison=(i == 0))
avg_napfd.append(res)
return avg_napfd
## State is represented by [Duration Gro-up , Time Slice , Last Histlen Verdicts]
## Hence the state size is 2 + hislen
## In future Other Important features may also be added to represent the state e.g the priority etc. and the state size has to be changed then.
state_size = 2 + args.histlen
## Preprocess Function are Used for State Representation.They are used for representing a test case as a state .
## Preprocess Continuous gives continuous values to the time group and the duration slice.
## Preprocess Discrete gives discrete values [0,1,2] to the time group and the duration state.
preprocess_function = preprocess_discrete
## Initializing the agent
if args.agent == 'tableau':
agent = agents.TableauAgent(learning_rate=args.learning_rate,
state_size=state_size,
action_size=args.actions,
epsilon=args.epsilon,
histlen=args.histlen)
## If the action size is 1 , we use MLPClassifier
## for action size as 2, we use MLPRegressor
elif args.agent == 'network':
if args.reward in ('binary'):
action_size = 1
else:
action_size = 2
agent = agents.NetworkAgent(state_size=state_size,
action_size=action_size,
hidden_size=args.hiddennet,
histlen=args.histlen)