def test(agents_nb, continued = False, erase = True, steps_nb = 100, probability=0.33, test_duration = 400, alpha=0.3, gamma=0.8, epsilon=0.8, training_params={'cycle_nb': 1, 'prob_step': 10, 'game_duration': 20}):
agent_list = [Agent() for i in range(agents_nb)]
old_results = []
if continued and agent_exists("stat2d"+str(agents_nb)):
training_params, probability, agents_nb, old_results, agent_list = load_test("stat2d"+str(agents_nb), agent_list)
hits = old_results + [[0 for i in range(agents_nb)] for j in range(steps_nb)]
game = Game(probability, 5)
for agent_id in range(agents_nb):
agent = agent_list[agent_id]
for step in range(steps_nb):
game.reset()
hits[len(hits) - steps_nb + step][agent_id] = play_game(agent, game, test_duration)
agent = train(agent = agent, training_params=training_params, learn_rate=alpha, discount_rate=gamma, policy=epsilon, show_prints=False)
print("Agent", agent_id+1)
if erase:
file = open ("saves/stat2d"+str(agents_nb)+".json", "w")
data = {}
for i in range(agents_nb):
data["data"+str(i)] = {str(k): v for k,v in agent_list[i].actions_value.items()}
to_write = json.dumps({
'data': data,
'options': {
'training_params': training_params,
'test_params': {
'probability': probability,
'agents_nb': agents_nb,
}
},
'old_results': hits
})
file.write(to_write)
file.close()
file = open ("saves/stat2d"+str(agents_nb)+".json", "r")
return hits
import gym
import monte_carlo
import q_learning
if __name__ == '__main__':
env = gym.make("FrozenLakeEasy-v0")
monte_carlo.train(env=env)
q_learning.train(env=env)
agent_list = [Agent() for i in range(10)]
game = Game(0, 5)
for agent_id in range(len(agent_list)):
agent = agent_list[agent_id]
for cycle in range(cycle_nb):
for i in range(prob_step):
game.reset()
game.probability = float(i) / prob_step
nb_hits = play_game(agent, game, 100)
hits[cycle, i] += nb_hits
max_hits[cycle, i] = max(max_hits[cycle, i], nb_hits)
min_hits[cycle, i] = min(min_hits[cycle, i], nb_hits)
agent = train(agent=agent,
training_params={
'cycle_nb': cycles_per_cycle,
'prob_step': 10,
'game_duration': 10
},
show_prints=False)
print("Cycle", cycle + 1)
hits /= len(agent_list)
ax = plt.figure().add_subplot(111, projection="3d")
X, Y = np.meshgrid(np.arange(0, 1, prob_step / 100),
np.arange(0, cycle_nb * cycles_per_cycle, cycles_per_cycle))
ax.plot_surface(X, Y, hits, rstride=1, cstride=1, cmap='summer')
ax.set_xlabel("Shoot probability")
ax.set_ylabel("Cycles")
from q_learning import load_agent, agent_exists, Agent, train
from ui import start_simulation
if agent_exists('save_file'):
agent, _ = load_agent('save_file')
else:
agent = Agent()
want_to_train = input('Train ? (True/False) ')
if want_to_train == str(True):
agent = train(agent, 'save_file')
def choose_action(state):
global agent
return agent.choose_best_action(state)
start_simulation(choose_action)