env = make_multi_env(env_init=env_init, nb_env=int(args.nb_env))
tp = TrainingParam()
# NN training
tp.lr = 1e-5
tp.lr_decay_steps = 300000
tp.minibatch_size = 32 * int(args.nb_env)
tp.update_freq = tp.minibatch_size / 2
# limit the number of time steps played per scenarios
tp.step_increase_nb_iter = None # None to deactivate it
tp.min_iter = None
tp.update_nb_iter = None # once 100 scenarios are solved, increase of "step_increase_nb_iter"
# oversampling hard scenarios
tp.oversampling_rate = None # None to deactivate it
# experience replay
tp.buffer_size = 1000000
# just observe the data for a while
tp.min_observe = None # int(10000)
# e greedy
tp.min_observation = 128
tp.initial_epsilon = 0.2
tp.final_epsilon = 1./(288.)
tp.step_for_final_epsilon = int(1e5)
# TODO add the "i dont do anything for a few time steps at the beginning of the training"
# don't start always at the same hour (if not None) otherwise random sampling, see docs
tp = TrainingParam()
# NN training
tp.lr = 1e-4
tp.lr_decay_steps = 30000
tp.minibatch_size = 256
tp.update_freq = 128
# limit the number of time steps played per scenarios
tp.step_increase_nb_iter = 2
tp.min_iter = 10
tp.update_nb_iter(2)
# oversampling hard scenarios
tp.oversampling_rate = 3
# experience replay
tp.buffer_size = 1000000
# e greedy
tp.min_observation = 10000
tp.initial_epsilon = 0.4
tp.final_epsilon = 1. / (2 * 7 * 288.)
tp.step_for_final_epsilon = int(1e5)
# don't start always at the same hour (if not None) otherwise random sampling, see docs
tp.random_sample_datetime_start = None
# saving, logging etc.
tp.save_model_each = 10000