def get_players(state):
blue_player = Entity()
blue_player.x = state[0]
blue_player.y = state[1]
red_player = Entity()
red_player.x = state[2]
red_player.y = state[3]
return blue_player, red_player
def can_escape_by_one_step(point1, point2):
env = Environment()
env.red_player = Entity()
env.red_player.x = point1[0]
env.red_player.y = point1[1]
env.blue_player = Entity()
env.blue_player.x = point2[0]
env.blue_player.y = point2[1]
win_stat = env.compute_terminal()
if win_stat == WinEnum.Blue:
return True
return False
def learn_decision_maker(decision_maker, n_samples = 20, save_to_file=False):
env = Environment()
blue_decision_maker = decision_maker # the agent we want to learn its behaviour policy
red_decision_maker = decision_maker # the imaginary agent we want to learn its behaviour policy
env.blue_player = Entity(blue_decision_maker)
env.red_player = Entity(red_decision_maker)
policy_counts = {}
# go over each states = (blue position, red position)
for state in state_generator():
if is_terminal_state(state):
continue
# set position of the players in the environment
set_env_state(env, state)
# get observation
observation_for_blue: State = env.get_observation_for_blue()
# the agents are not deterministic, so we want to find the distribution p(a|s)
for i_samp in range(n_samples):
# get the action chosen by each player
action_blue = blue_decision_maker.get_action(observation_for_blue)
a = action_blue - 1 # change to 0-based index
update_pol_cnts(state, a, policy_counts)
# end for
# end for
print('Finished learning the enemy')
if save_to_file:
with open(f'learned_{agent_name}_enemy', 'wb') as myfile:
pickle.dump([agent_name, policy_counts, n_samples], myfile)
return policy_counts
import pickle
import numpy as np
import json
from Arena.Environment import Environment
from Arena.Entity import Entity
from RafaelPlayer.RafaelDecisionMaker import RafaelDecisionMaker
from Arena.constants import WIN_REWARD, MOVE_PENALTY, MAX_STEPS_PER_EPISODE, HARD_AGENT
from misharon_utils import state_action_generator, get_Q_vals, set_env_state, is_terminal_state, derive_greedy_policy
from misharon_learn_the_enemy import n_actions
#------------------------------------------------------------------------------------------------------------~
# define dummy players, just so we can use the class functions
dummy_blue = Entity(RafaelDecisionMaker(HARD_AGENT))
dummy_red = Entity(RafaelDecisionMaker(HARD_AGENT))
env = Environment()
env.blue_player = dummy_blue
env.red_player = dummy_red
#------------------------------------------------------------------------------------------------------------~
# ------------------------------------------------------------------------------------------------------------~
def get_reward(env, state):
set_env_state(env, state)
reward_blue, reward_red = env.handle_reward()
return reward_blue
if __name__ == '__main__':
env = Environment(IS_TRAINING)
print("Starting Blue player")
blue_decision_maker = DQNAgent_keras.DQNAgent_keras()
#blue_decision_maker = DQNAgent_keras.DQNAgent_keras(UPDATE_CONTEXT=True, path_model_to_load='conv1(6_6_1_256)_conv2(4_4_256_128)_conv3(3_3_128_128)_flatten_fc__blue_202001_ 0.95max_ -0.04avg_ -3.10min__1620558885.model')
print("Starting red player")
### Red Decision Maker
red_decision_maker = Greedy_player.Greedy_player()
env.blue_player = Entity(blue_decision_maker)
env.red_player = Entity(red_decision_maker)
print_start_of_game_info(blue_decision_maker, red_decision_maker)
NUM_OF_EPISODES = env.NUMBER_OF_EPISODES
for episode in tqdm(range(1, NUM_OF_EPISODES + 1), ascii=True, unit='episodes'):
EVALUATE = evaluate(episode)
current_episode = Episode(episode, EVALUATE, show_always=False if IS_TRAINING else True)
# set new start position for the players
env.reset_game(episode)
# get observation
observation_for_blue_s0: State = env.get_observation_for_blue()
action_blue = -1