def get_features_values_by_competition(graph):
iterations = run.feature_iterations
features = np.zeros((iterations * 1600, 7))
counter = 0
for iteration in range(iterations):
g = graph.copy()
while len(g.graph['free']) > 1:
feature = tools.get_feature(g)
features[counter][:] = feature
counter += 1
# first player action
action_choice1 = np.random.randint(3)
if action_choice1 == 0:
seed1 = actions.action_degree(g)
elif action_choice1 == 1:
seed1 = actions.action_weight(g)
elif action_choice1 == 2:
seed1 = actions.action_blocking(g, 1)
# illegal action
if seed1 == -1:
print("Illegal action")
action_choice1 = 0
seed1 = actions.action_degree(g)
tools.activate_node(g, seed1, 1)
# second player action
action_choice2 = np.random.randint(3)
if action_choice2 == 0:
seed2 = actions.action_degree(g)
elif action_choice2 == 1:
seed2 = actions.action_weight(g)
elif action_choice2 == 2:
seed2 = actions.action_blocking(g, 2)
# illegal action
if seed2 == -1:
print("Illegal action")
action_choice2 = 0
seed2 = actions.action_degree(g)
tools.activate_node(g, seed2, 2)
a1, a2 = tools.diffuse(g)
for n in a1:
tools.activate_node(g, n, 1)
for n in a2:
tools.activate_node(g, n, 2)
return features[0:counter, :]
def train(graph, feature_lmh_ranges):
alpha = 0.5
eps = 0.8
gamma = 0.98
d = 0.998
opponent_action = run.opponent
q_table = np.zeros((2187, 3))
for iteration in tqdm(range(run.train_n_rounds)):
g = graph.copy()
state = tools.get_state(g, feature_lmh_ranges)
for t in range(21):
# first player action
if np.random.random() < eps:
action_choice1 = np.random.randint(3)
else:
action_choice1 = np.argmax(q_table[state, :])
if action_choice1 == 0:
seed1 = actions.action_degree(g)
elif action_choice1 == 1:
seed1 = actions.action_weight(g)
elif action_choice1 == 2:
seed1 = actions.action_blocking(g, 1)
# elif action_choice1 == 3:
# seed1 = actions.action_last(g)
# elif action_choice1 == 4:
# seed1 = actions.action_min_degree(g)
# illegal action
if seed1 == -1:
print("Illegal action")
action_choice1 = 0
seed1 = actions.action_degree(g)
tools.activate_node(g, seed1, 1)
# second player action
action_choice2 = opponent_action
if action_choice2 == 0:
seed2 = actions.action_degree(g)
elif action_choice2 == 1:
seed2 = actions.action_weight(g)
elif action_choice2 == 2:
seed2 = actions.action_blocking(g, 2)
# illegal action
if seed2 == -1:
print("Illegal action")
action_choice2 = 0
seed2 = actions.action_degree(g)
tools.activate_node(g, seed2, 2)
a1, a2 = tools.diffuse(g)
for n in a1:
tools.activate_node(g, n, 1)
for n in a2:
tools.activate_node(g, n, 2)
next_state = tools.get_state(g, feature_lmh_ranges)
# delayed reward
r = 0
if t % 6 == 0:
r = len(g.graph['1']) - len(g.graph['2'])
# immediate reward
# r = len(g.graph['1']) - len(g.graph['2'])
q_table[state, action_choice1] = (
1 - alpha) * q_table[state, action_choice1] + alpha * (
r + gamma * max(q_table[next_state, :]))
state = next_state
alpha = d * alpha
eps = 0.9 - (0.6 * iteration / 500)
return q_table
def compete(graph: nx.Graph, q_table):
opponent_action = 0
counter = 0
max_iteration = 20000000
un_seen = 0
while len(graph.graph['free']) > 1 and counter < max_iteration:
state = tools.get_state(graph)
# first player action
action_choice1 = np.argmax(q_table[state, :])
if np.sum(q_table[state, :]) == 0:
un_seen += 1
action_choice1 = np.random.randint(5)
if action_choice1 == 0:
seed1 = actions.action_degree(graph)
elif action_choice1 == 1:
seed1 = actions.action_weight(graph)
elif action_choice1 == 2:
seed1 = actions.action_blocking(graph, 1)
elif action_choice1 == 3:
seed1 = actions.action_last(graph)
elif action_choice1 == 4:
seed1 = actions.action_min_degree(graph)
# illegal action
if seed1 == -1:
print("illegal 1")
print(action_choice1)
action_choice1 = 0
seed1 = actions.action_degree(graph)
tools.activate_node(graph, seed1, 1)
# second player action
action_choice2 = opponent_action
if action_choice2 == 0:
seed2 = actions.action_degree(graph)
elif action_choice2 == 1:
seed2 = actions.action_weight(graph)
elif action_choice2 == 2:
seed2 = actions.action_blocking(graph, 2)
# illegal action
if seed2 == -1:
print("illegal 2")
print(action_choice2)
action_choice2 = 0
seed2 = actions.action_degree(graph)
tools.activate_node(graph, seed2, 2)
a1, a2 = tools.diffuse(graph)
for n in a1:
tools.activate_node(graph, n, 1)
for n in a2:
tools.activate_node(graph, n, 2)
counter += 1
score = len(graph.graph['1']) - len(graph.graph['2'])
print("counter: " + str(counter))
print("unseen: " + str(un_seen))
print("score: " + str(score))
result = {"counter": counter, "un_seen": un_seen, "score": score}
return result