def testing(self):
training = True
while training:
tree = Tree(self.curr_network)
tree.black_root = Node(
np.array([
0, 0, 0, 0, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0,
0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 6, 7, 0, 0, 0, 0, 0, 0])
)
tree.black_root.build_actions()
tree.set_probabilities(tree.black_root)
tree.white_root = Node(np.array([
0, 0, 0, 0, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 6, 7, 0, 0, 0, 0, 0, 0, 0,
0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 5, 6, 7, 0, 0, 0, 1, 0, 0])
)
tree.white_root.build_actions()
tree.set_probabilities(tree.white_root)
self.env.set_env(tree.black_root.state)
done = False
while not done:
tree.noise = True
for counter in range(0, 800):
tree.env.set_env(tree.black_root.state)
tree.test_simulations()
black_action_edge = max(tree.black_root.edges,key=attrgetter('visits'))
white_action_edge = max(tree.white_root.edges,key=attrgetter('visits'))
black_mc_prob = self.build_mc_prob(tree.black_root)
print([edge.visits for edge in tree.black_root.edges])
white_mc_prob = self.build_mc_prob(tree.white_root)
env_action = self.env.step(black_action_edge.action, white_action_edge.action)
black_score, white_score = self.env.get_player_scores()
black_state, white_state = self.build_states(tree.black_root.state, black_action_edge,
white_action_edge, env_action, black_score, white_score)
black_root, white_root = self.find_roots(black_state, white_state, black_action_edge, white_action_edge)
tree.black_root = black_root
tree.white_root = white_root
done = self.env.check_status()
if done:
print(self.env.eval_game())
def eval_net(self):
updated_net = 0
old_net = 0
draw =0
for i in range (0,100):
tree = Tree(self.learning_network)
tree.init_tree()
old_tree = Tree(self.curr_network)
old_tree.init_tree()
eval_env = Environment()
done = False
while not done:
for x in range (0,400):
tree.env.set_env(tree.black_root.state)
old_tree.env.set_env(old_tree.black_root.state)
tree.double_agent_simulation()
old_tree.double_agent_simulation()
action_edge = max(tree.black_root.edges,key=attrgetter('visits'))
action = action_edge.action
old_action_edge = max(old_tree.white_root.edges,key=attrgetter('visits'))
old_action = old_action_edge.action
env_action = eval_env.step(action,old_action)
new,old =eval_env.get_player_scores()
black_state,white_state = self.build_states(tree.black_root.state,action_edge,old_action_edge,env_action,new,old)
learning_black_root = Node(black_state)
learning_black_root.build_actions()
self.set_probabilities(learning_black_root,self.learning_network)
learning_white_root = Node(white_state)
learning_white_root.build_actions()
self.set_probabilities(learning_white_root,self.learning_network)
tree.black_root = learning_black_root
tree.white_root = learning_white_root
curr_black_root = Node(black_state)
curr_black_root.build_actions()
self.set_probabilities(curr_black_root, self.curr_network)
curr_white_root = Node(white_state)
curr_white_root.build_actions()
self.set_probabilities(curr_white_root, self.curr_network)
old_tree.black_root = curr_black_root
old_tree.white_root = curr_white_root
done = eval_env.check_status()
winner = eval_env.eval_game()
if winner == 1 :
updated_net +=1
print("Net won this game " + str(updated_net) )
elif winner == 0:
draw +=1
print("Draw " + str(draw))
else:
old_net +=1
print("Old Net won. " + str(old_net))
print("New Network won " + str(updated_net) + " matches against the old one.")
print("Draws: "+str(draw))
if updated_net / (updated_net + old_net) >= 0.54:
torch.save(self.learning_network.state_dict(), self.dir)
self.curr_network.load_state_dict(torch.load(self.dir))
self.curr_network.eval()
print("---Network updated---")
else:
self.learning_network.load_state_dict(torch.load(self.dir))
self.learning_network.eval()
