def run_simulation(config_filename, graphfilename, n):
model = BaseModel(config_filename)
tcm = TrueCausalModel(model)
tcm.model.save_pgm_as_img(graphfilename)
intervention_vars = model.get_intervention_variables()
target_value = 1
target = {
"variable": model.get_target_variable(),
"value": target_value
}
local_data = dict()
for i in range(n):
idx_intervention_var = np.random.randint(len(intervention_vars))
action = (intervention_vars[idx_intervention_var], np.random.randint(2))
nature_response = tcm.action_simulator([action[0]], [action[1]])
for k in nature_response:
if not k in local_data:
local_data[k] = []
local_data[k].append(nature_response[k])
df = pd.DataFrame.from_dict(local_data)
model_from_data = generate_approx_model_from_graph(ebunch=model.get_ebunch(), nodes=model.nodes, df=df)
for cpd in model_from_data.get_cpds():
logging.info(cpd)
def basic_model_learning(base_path="results/disease-treatment-best-action", experiments=10, rounds=50, plot_id=""):
gt_ebunch = [("Reaction", "Lives"), ("Treatment", "Reaction"), ("Treatment", "Lives"), ("Disease", "Lives")]
DG = nx.DiGraph([("Reaction", "Lives"), ("Treatment", "Reaction"), ("Treatment", "Lives"), ("Disease", "Lives")])
# causal_order = list(nx.topological_sort(DG))
causal_order = []
# invalid_edges = [("Disease", "Treatment")]
invalid_edges = []
COMPLETE_MODEL = BaseModel('configs/model_parameters.json')
MODEL_REACTION = BaseModel('configs/model_parameters_reaction.json')
MODEL_LIVES = BaseModel('configs/model_parameters_lives.json')
nature = TrueCausalModel(COMPLETE_MODEL)
nature_reaction = TrueCausalModel(MODEL_REACTION)
nature_lives = TrueCausalModel(MODEL_LIVES)
nature_arr = [nature, nature_reaction, nature_lives]
variables = sorted(["Treatment", "Reaction", "Disease", "Lives"])
intervention_vars = COMPLETE_MODEL.get_intervention_variables()
target_value = 1
target = {
"variable": COMPLETE_MODEL.get_target_variable(),
"value" : target_value
}
g_truth = {e: 1 for e in DG.edges}
n_exploration_steps = 1
global_results = dict()
for i in range(experiments):
start_time = time.time()
base_experiment_filename = f"disease-treatment-run-{i}-rounds-{rounds}"
results_data = dict()
local_exp_results = dict()
data = explore_and_generate_data(nature, intervention_vars, n_steps=n_exploration_steps)
df = pd.DataFrame.from_dict(data)
connection_tables = create_pij(variables, causal_order, invalid_edges, use_causal_order=False)
adj_list = create_graph_from_beliefs_unknown_order(variables, connection_tables)
ebunch, nodes = adj_list_to_ebunch_and_nodes(adj_list)
approx_model = generate_approx_model_from_graph(ebunch, nodes, df)
unknown_model = BaseModel('configs/incomplete_params.json')
unknown_model.reset(approx_model, ebunch, nodes)
unknown_model.show_graph()
connection_probas, rewards = training(
variables, rounds, connection_tables, data, unknown_model, nature_arr, target)
for key in connection_probas:
if key not in global_results:
global_results[key] = []
global_results[key].append(connection_probas[key])
local_exp_results[key] = connection_probas[key]
results_data[f"gt_{i}"] = g_truth
results_data[f"beliefs_{i}"] = local_exp_results
results_data[f"training_time_{i}"] = time.time() - start_time
results_data[f"rewards_{i}"] = rewards
dict_filename = os.path.join(
base_path, "mats", base_experiment_filename + ".pickle")
with open(dict_filename, "wb") as handle:
pickle.dump(results_data, handle, protocol=pickle.HIGHEST_PROTOCOL)
labels = []
mean_vectors = []
std_dev_vectors = []
labels_correct_ones = []
mean_vectors_correct_ones = []
std_dev_vectors_correct_ones = []
labels_wrong_ones = []
mean_vectors_wrong_ones = []
std_dev_vectors_wrong_ones = []
for key in global_results:
labels += [key]
mean_vectors.append(np.mean(global_results[key], axis=0))
std_dev_vectors.append(np.std(global_results[key], axis=0))
if key in gt_ebunch:
labels_correct_ones.append(key)
mean_vectors_correct_ones.append(mean_vectors[-1])
std_dev_vectors_correct_ones.append(std_dev_vectors[-1])
else:
labels_wrong_ones.append(key)
mean_vectors_wrong_ones.append(mean_vectors[-1])
std_dev_vectors_wrong_ones.append(std_dev_vectors[-1])
print("{} {} {}".format(key, mean_vectors[-1][-1], std_dev_vectors[-1][-1]))
x_axis = np.arange(len(mean_vectors[0]))
plot_measures(x_axis, mean_vectors, std_dev_vectors, labels,
f"{base_path}/{plot_id}_connection_beliefs_exp_{experiments}_rounds_{rounds}_{intervention_vars}", outside_legend=True)
plot_measures(x_axis, mean_vectors_correct_ones, std_dev_vectors_correct_ones, labels_correct_ones,
f"{base_path}/{plot_id}_connection_beliefs_correct_exp_{experiments}_rounds_{rounds}_{intervention_vars}", outside_legend=True)
plot_measures(x_axis, mean_vectors_wrong_ones, std_dev_vectors_wrong_ones, labels_wrong_ones,
f"{base_path}/{plot_id}_connection_beliefs_wrong_exp_{experiments}_rounds_{rounds}_{intervention_vars}", outside_legend=True)
