import numpy as np
import matplotlib
matplotlib.use('TkAgg')
from lib.envs.bandit import BanditEnv
from lib.simulation import Experiment
from shared.policy import UCB
evaluation_seed = 1239
num_actions = 10
trials = 100
distribution = "normal"
env = BanditEnv(num_actions, distribution, evaluation_seed)
agent = UCB(num_actions)
experiment = Experiment(env, agent)
experiment.run_bandit(trials)
"maze": SimpleRoomsEnv(),
"grid": SimpleRoomsEnv()
}
return switcher.get(argument)
env = get_env(env_string)
if agent_string.startswith('q'):
print("Running Q Learning on {} environment for {} epochs".format(
env_string, num_iter))
agent = QLearningAgent(range(env.action_space.n),
epsilon=epsilon,
alpha=alpha,
decay_every=decay)
experiment = Experiment(env, agent)
experiment.run_qlearning(num_iter, interactive)
#print("Running Q Learning")
elif agent_string.startswith('s'):
print("Running SARSA on {} environment for {} epochs".format(
env_string, num_iter))
agent = SarsaAgent(range(env.action_space.n),
epsilon=epsilon,
alpha=alpha,
decay_every=decay)
experiment = Experiment(env, agent)
experiment.run_sarsa(num_iter, interactive)
#print("Running SARSA")
else:
print("Invalid Agent argument")
max_value_indices.append(idx)
return np.random.choice(max_value_indices)
def learn(self, state1, action1, reward, state2, stop):
"""
SARSA Update
Q(s,a) <- Q(s,a) + alpha * (reward + gamma * Q(s',a') - Q(s,a))
or
Q(s,a) <- Q(s,a) + alpha * (td_target - Q(s,a))
or
Q(s,a) <- Q(s,a) + alpha * td_delta
"""
self._Q_table[state1][action1] = self._Q_table[state1][action1] + \
self._alpha * (
reward + self._gamma * max(self._Q_table[state2]) - \
self._Q_table[state1][action1]
)
interactive = True
env = WindyGridworldEnv()
agent = QLearningAgent(
range(env.action_space.n),
env.S,
)
experiment = Experiment(env, agent)
experiment.run_sarsa(10, interactive)
import numpy as np
import matplotlib
matplotlib.use('TkAgg')
from lib.envs.simple_rooms import SimpleRoomsEnv
from lib.simulation import Experiment
from shared.agent import RandomAgent
interactive = True
max_number_of_episodes = 5
env = SimpleRoomsEnv()
agent = RandomAgent(range(env.action_space.n))
experiment = Experiment(env, agent)
experiment.run_agent(max_number_of_episodes, interactive)
## TODO 3
## Implement the q-learning update here
"""
Q-learning Update:
Q(s,a) <- Q(s,a) + alpha * (reward + gamma * max(Q(s') - Q(s,a))
or
Q(s,a) <- Q(s,a) + alpha * (td_target - Q(s,a))
or
Q(s,a) <- Q(s,a) + alpha * td_delta
"""
interactive = True
env = SimpleRoomsEnv()
agent = QLearningAgent(range(env.action_space.n))
experiment = Experiment(env, agent)
experiment.run_qlearning(10, interactive)
interactive = False
env = SimpleRoomsEnv()
agent = QLearningAgent(range(env.action_space.n))
experiment = Experiment(env, agent)
experiment.run_qlearning(50, interactive)
interactive = True
env = CliffWalkingEnv()
agent = QLearningAgent(range(env.action_space.n))
experiment = Experiment(env, agent)
experiment.run_qlearning(10, interactive)
interactive = False