def run_simulations(args, save_result, local_mode, init_ray=True):
if init_ray:
start_ray(local_mode)
if save_result:
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
set_random_seed(args.seed)
n_reps = args.n_reps
alg_param_grid = get_grid(args.param_grid_def)
n_grid = alg_param_grid.shape[0]
config_grid_vals = get_grid(args.config_grid_def)
n_config_grid = len(config_grid_vals)
planing_loss = np.zeros((n_reps, n_config_grid, n_grid))
info_dict = {}
# ----- Run simulation in parrnell process---------------------------------------------#
loss_rep_id_lst = []
for i_rep in range(n_reps):
# returns objects ids:
args_r = deepcopy(args)
planing_loss_rep_id = run_rep.remote(i_rep, alg_param_grid,
args_r.config_grid_def, args_r)
loss_rep_id_lst.append(planing_loss_rep_id)
# end for i_rep
# ----- get the results --------------------------------------------#
for i_rep in range(n_reps):
loss_rep = ray.get(loss_rep_id_lst[i_rep])
if i_rep % max(n_reps // 100, 1) == 0:
time_str = time.strftime(
"%H hours, %M minutes and %S seconds",
time.gmtime(timeit.default_timer() - start_time))
write_to_log(
'Finished: {} out of {} reps, time: {}'.format(
i_rep + 1, n_reps, time_str), args)
# end if
planing_loss[i_rep] = loss_rep
info_dict = {
'planing_loss_avg': planing_loss.mean(axis=0),
'planing_loss_std': planing_loss.std(axis=0),
'alg_param_grid': alg_param_grid,
'n_reps_finished': i_rep + 1
}
if save_result:
save_run_data(args, info_dict, verbose=0)
# end if
# end for i_rep
if save_result:
save_run_data(args, info_dict)
stop_time = timeit.default_timer()
write_to_log(
'Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(stop_time - start_time)), args, save_result)
return info_dict
def run_rep(i_rep, alg_param_grid, config_grid_def, args_r):
set_random_seed(args_r.seed + i_rep)
config_grid_vals = get_grid(config_grid_def)
n_config_grid = len(config_grid_vals)
n_grid = len(alg_param_grid)
# runs a single repetition of the experiment
loss_rep = np.zeros((n_config_grid, n_grid))
gammaEval = args_r.gammaEval
config_type = config_grid_def['type']
# grid of number of trajectories to generate
for i_config, config_val in enumerate(config_grid_vals):
n_traj = args_r.n_trajectories
if config_type == 'n_trajectories':
n_traj = config_val
elif config_type == 'states_actions_TV_dist_from_uniform':
args_r.train_sampling_def = {
'type': 'Generative',
'states_TV_dist_from_uniform': config_val,
'actions_TV_dist_from_uniform': config_val
}
elif config_type == 'chain_mix_time':
args_r.train_sampling_def = {
'type': 'chain_mix_time',
'mix_time': config_val
}
elif config_type == 'n_episodes':
args_r.n_episodes = config_val
elif config_type == 'None':
pass
else:
raise AssertionError
# Generate MDP:
M = MDP(args_r)
# Optimal policy for the MDP:
pi_opt, V_opt, Q_opt = PolicyIteration(M, gammaEval)
# grid of regularization param
for i_grid, alg_param in enumerate(alg_param_grid):
gamma_guidance, l1_factor, l2_factor = get_regularization_params(
args_r, alg_param, args_r.param_grid_def['type'])
# run the learning episodes:
pi_t = run_learning_method(args_r, M, n_traj, gamma_guidance,
l2_factor, l1_factor)
# Evaluate performance of learned policy:
V_t, _ = PolicyEvaluation(M, pi_t, gammaEval)
loss_rep[i_config, i_grid] = (np.abs(V_opt - V_t)).mean()
# end for grid
# end for i_config
return loss_rep
def run_simulation(args, hyper_grid_vals, loss, reg_grids, local_mode):
start_ray(local_mode)
write_to_log('local_mode == {}'.format(local_mode), args)
SetMrpArgs(args)
start_time = timeit.default_timer()
set_random_seed(args.seed)
reg_types = args.reg_types
n_hyper_grid = len(hyper_grid_vals)
n_reps = args.n_reps
results_dict = dict()
write_to_log('***** Starting {} reps'.format(n_reps), args)
for i_rep in range(n_reps):
for i_hyper_grid, hyper_grid_val in enumerate(hyper_grid_vals):
args_run = deepcopy(args)
set_hyper_param(args_run, hyper_grid_val)
# send jobs:
out_ids = {reg_type: [None for _ in range(len(reg_grids[reg_type]))] for reg_type in reg_types}
for reg_type in reg_types:
for i_reg_pram, reg_param in enumerate(reg_grids[reg_type]):
# ray put
if np.isnan(loss[reg_type][i_hyper_grid, i_reg_pram, i_rep]):
out_ids[reg_type][i_reg_pram] = run_exp.remote(i_rep, args_run, reg_type, reg_param)
# end if
# end for i_reg_pram
# end for reg_type
# Gather results:
for reg_type in reg_types:
for i_reg_pram, reg_param in enumerate(reg_grids[reg_type]):
# ray get
if out_ids[reg_type][i_reg_pram] is not None:
out = ray.get(out_ids[reg_type][i_reg_pram])
loss[reg_type][i_hyper_grid, i_reg_pram, i_rep] = out
# end if
# end for i_reg_pram
# end for reg_type
# end for i_hyper_grid
# Save results so far
results_dict = {'hyper_grid_vals': hyper_grid_vals, 'loss': loss, 'reg_grids': reg_grids, 'n_reps_finished': i_rep + 1}
save_run_data(args, results_dict, verbose=0)
time_str = time.strftime("%H hours, %M minutes and %S seconds", time.gmtime(timeit.default_timer() - start_time))
write_to_log('Finished: {} out of {} reps, time: {}'.format(i_rep + 1, n_reps, time_str), args)
# end for i_rep
stop_time = timeit.default_timer()
write_to_log('Total runtime: ' + time.strftime("%H hours, %M minutes and %S seconds", time.gmtime(stop_time - start_time)), args)
return results_dict
def run_simulations(args, save_result, local_mode):
args_def = deepcopy(args)
start_ray(local_mode)
if save_result:
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
set_random_seed(args.seed)
n_reps = args.n_reps
param_val_grid = get_grid(args.param_grid_def)
n_grid = param_val_grid.shape[0]
config_grid = get_grid(args.config_grid_def)
n_configs = len(config_grid)
args.n_configs = n_configs
loss_mat = np.zeros((n_reps, n_configs, n_grid))
# ----- Run simulation in parrnell process---------------------------------------------#
loss_rep_id_lst = []
for i_rep in range(n_reps):
# returns objects ids:
loss_mat_rep_id = run_rep.remote(i_rep, param_val_grid, config_grid,
args)
loss_rep_id_lst.append(loss_mat_rep_id)
# ----- get the results --------------------------------------------#
for i_rep in range(n_reps):
loss_rep = ray.get(loss_rep_id_lst[i_rep])
write_to_log('Finished: {} out of {} reps'.format(i_rep + 1, n_reps),
args)
loss_mat[i_rep] = loss_rep
# end for i_rep
info_dict = {
'loss_avg': loss_mat.mean(axis=0),
'loss_std': loss_mat.std(axis=0),
'param_val_grid': param_val_grid,
'config_grid': config_grid
}
if save_result:
save_run_data(args, info_dict)
stop_time = timeit.default_timer()
write_to_log(
'Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(stop_time - start_time)), args)
write_to_log(
['-' * 10 + 'Defined args: ',
pretty_print_args(args_def), '-' * 20], args)
return info_dict
def run_simulations(args, local_mode):
start_ray(local_mode)
create_result_dir(args)
write_to_log('local_mode == {}'.format(local_mode), args)
start_time = timeit.default_timer()
create_result_dir(args)
set_random_seed(args.seed)
l2_grid = get_grid(args.l2_grid_def)
gam_grid = get_grid(args.gam_grid_def)
write_to_log('gamma_grid == {}'.format(gam_grid), args)
write_to_log('l2_grid == {}'.format(l2_grid), args)
grid_shape = (len(l2_grid), len(gam_grid))
loss_avg = np.zeros(grid_shape)
loss_std = np.zeros(grid_shape)
run_idx = 0
for i0 in range(grid_shape[0]):
for i1 in range(grid_shape[1]):
args_run = deepcopy(args)
args_run.param_grid_def = {
'type': 'L2_factor',
'spacing': 'list',
'list': [l2_grid[i0]]
}
args_run.default_gamma = gam_grid[i1]
info_dict = run_main_control(args_run,
save_result=False,
plot=False,
init_ray=False)
loss_avg[i0, i1] = info_dict['planing_loss_avg'][0]
loss_std[i0, i1] = info_dict['planing_loss_std'][0]
run_idx += 1
print("Finished {}/{}".format(run_idx, loss_avg.size))
# end for
# end for
grid_results_dict = {
'l2_grid': l2_grid,
'gam_grid': gam_grid,
'loss_avg': loss_avg,
'loss_std': loss_std
}
save_run_data(args, grid_results_dict)
stop_time = timeit.default_timer()
write_to_log(
'Total runtime: ' +
time.strftime("%H hours, %M minutes and %S seconds",
time.gmtime(stop_time - start_time)), args)
return grid_results_dict
def run_exp(i_rep, args_run, reg_type, reg_param): # set seed set_random_seed(args_run.seed + i_rep) # Generate MDP and sampling distribution (with specified uniformity) M = MRP(args_run) gammaEval = args_run.gammaEval # set regularisation parameters gamma_guidance, l2_TD, l2_fp, l2_proj = get_regularization_params(args_run, reg_param, reg_type) # Generate data: data = M.SampleDataMrp(args_run) V_est, V_true = run_value_estimation_method(data, M, args_run, gamma_guidance, l2_proj, l2_fp, l2_TD) loss_type = args_run.evaluation_loss_type pi = None eval_loss = evaluate_value_estimation(loss_type, V_true, V_est, M, pi, gammaEval, gamma_guidance) return eval_loss