# init two models
model1 = DeepQNetwork(env, predator, "predator")
model2 = DeepQNetwork(env, prey, "prey")
# load trained model
model1.load("data/pursuit_model")
model2.load("data/pursuit_model")
done = False
step_ct = 0
print("nums: %d vs %d" % (env.get_num(predator), env.get_num(prey)))
while not done:
# take actions for deers
obs_1 = env.get_observation(predator)
ids_1 = env.get_agent_id(predator)
acts_1 = model1.infer_action(obs_1, ids_1)
env.set_action(predator, acts_1)
# take actions for tigers
obs_2 = env.get_observation(prey)
ids_2 = env.get_agent_id(prey)
acts_2 = model2.infer_action(obs_2, ids_1)
env.set_action(prey, acts_2)
# simulate one step
done = env.step()
# render
env.render()
# get reward
# init two models
model1 = DeepQNetwork(env, predator, "predator")
model2 = DeepQNetwork(env, prey, "prey")
# load trained model
model1.load("data/pursuit_model")
model2.load("data/pursuit_model")
done = False
step_ct = 0
print("nums: %d vs %d" % (env.get_num(predator), env.get_num(prey)))
while not done:
# take actions for deers
obs_1 = env.get_observation(predator)
ids_1 = env.get_agent_id(predator)
acts_1 = model1.infer_action(obs_1, ids_1)
env.set_action(predator, acts_1)
# take actions for tigers
obs_2 = env.get_observation(prey)
ids_2 = env.get_agent_id(prey)
acts_2 = model2.infer_action(obs_2, ids_1)
env.set_action(prey, acts_2)
# simulate one step
done = env.step()
# render
env.render()
# get reward
for j in range(n_agent):
ob.append(np.asarray([flat_ob[j]]))
ob.append(np.asarray([adj[j]]))
ob.append(np.asarray([vec]))
acts = model.predict(ob)
action[i] = np.zeros(n_agent, dtype=np.int32)
for j in range(n_agent):
if np.random.rand() < alpha:
action[i][j] = random.randrange(n_actions)
else:
action[i][j] = np.argmax(acts[j])
env.set_action(handles[i], action[i])
obs[1] = env.get_observation(handles[1])
ids[1] = env.get_agent_id(handles[1])
acts = tf_model.infer_action(obs[1], ids[1], 'e_greedy')
env.set_action(handles[1], acts)
done = env.step()
next_obs = env.get_observation(handles[0])
flat_next_obs = observation(next_obs[0], next_obs[1])
rewards = env.get_reward(handles[0])
score += sum(rewards)
if steps % 3 == 0:
buff.add(flat_ob, action[0], flat_next_obs, rewards, done, adj)
if (i_episode - 1) % 10 == 0:
env.render()
if max_steps == steps:
print(dead[0], end='\t')
print(dead[1], end='\t')
# tf.reset_default_graph()
model9 = DeepQNetwork(env, army9, "battle-l", memory_size=2**10)
#tf.reset_default_graph()
model10 = DeepQNetwork(env, army10, "battle-r", memory_size=2**10)
model9.load("save_model", 14)
model10.load("save_model", 14)
done = False
step_ct = 0
print("nums: %d vs %d" % (env.get_num(army1), env.get_num(army2)))
while not done:
# take actions for army1
obs_1 = env.get_observation(army1)
ids_1 = env.get_agent_id(army1)
acts_1 = model1.infer_action(obs_1, ids_1)
env.set_action(army1, acts_1)
# take actions for army2
obs_2 = env.get_observation(army2)
ids_2 = env.get_agent_id(army2)
acts_2 = model2.infer_action(obs_2, ids_2)
env.set_action(army2, acts_2)
# take actions for army3
obs_3 = env.get_observation(army3)
ids_3 = env.get_agent_id(army3)
acts_3 = model3.infer_action(obs_3, ids_3)
env.set_action(army3, acts_3)
# take actions for army4