class TestMdp(unittest.TestCase):
def setUp(self):
self.mdp = MarkovDecisionProcess()
self.mdp.reset()
self.mdp.update(Percept(0, 2, 1, 1, False))
self.mdp.update(Percept(0, 2, 4, 2, False))
def test_rewards(self):
self.assertEqual(self.mdp.rewards[0, 2, 1], 1)
def test_nsa(self):
self.assertEqual(self.mdp.nsa[0, 2], 2)
def test_ntsa(self):
self.assertEqual(self.mdp.ntsa[0, 2, 1], 1)
def test_ptsa(self):
self.assertEqual(self.mdp.ptsa[0, 2, 1], 0.5)
class Main:
"""
Solely for testing purposes. Do not use.
"""
mdp = MarkovDecisionProcess(0, range(16), range(4))
strategy = Qlearning(mdp, 1, 0.1, 1.0, 0.01)
algorithm = Agent(strategy, 'FrozenLake-v0')
algorithm.learn(1000)
"""
Use code below to visualise evolution of epsilon
When not needed comment out as well as code in learningstrategy.py
"""
"""
class FlGame(GridLayout):
"""
grid layout visualisation of Frozen Lake game
the learning algorithm is initialised within so make adjustments here
"""
n_episodes = 50000
mdp = MarkovDecisionProcess(0, range(16), range(4))
# strategy = Qlearning(mdp, 0.1, 0.001, 0.9, 1.0, 0.01)
# strategy = NstepQlearning(mdp, 0.5, 0.001, 0.6, 4, 1.0, 0.01)
# strategy = MonteCarlo(mdp, 0.1, 0.001, 0.9, 1.0, 0.01)
strategy = ValueIteration(mdp, 0.8, 0.001, 0.90, 0.9, 1.0, 0.01)
algorithm = Agent(strategy, 'FrozenLake-v0')
state0 = ObjectProperty(None)
state1 = ObjectProperty(None)
state2 = ObjectProperty(None)
state3 = ObjectProperty(None)
state4 = ObjectProperty(None)
state5 = ObjectProperty(None)
state6 = ObjectProperty(None)
state7 = ObjectProperty(None)
state8 = ObjectProperty(None)
state9 = ObjectProperty(None)
state10 = ObjectProperty(None)
state11 = ObjectProperty(None)
state12 = ObjectProperty(None)
state13 = ObjectProperty(None)
state14 = ObjectProperty(None)
state15 = ObjectProperty(None)
def update(self, dt):
if self.algorithm.strategy.episode_count < self.n_episodes:
self.algorithm.learn(1)
if self.algorithm.strategy.episode_count % 100 == 0:
print(self.algorithm.strategy.episode_count)
actions = self.algorithm.visualisationList
self.state0.text = actions[0]
self.state1.text = actions[1]
self.state2.text = actions[2]
self.state3.text = actions[3]
self.state4.text = actions[4]
self.state6.text = actions[6]
self.state8.text = actions[8]
self.state9.text = actions[9]
self.state10.text = actions[10]
self.state13.text = actions[13]
self.state14.text = actions[14]
def setUp(self):
self.mdp = MarkovDecisionProcess()
self.value = ValueIteration(self.mdp, 0.8, 0.001, 0.90, 0.9, 1.0, 0.01)
self.value.mdp.reset()
self.value.evaluate(Percept(0, 2, 1, 1, False))
self.value.evaluate(Percept(0, 2, 4, 2, False))
import cProfile
from scripts.markovDecisionProcess import MarkovDecisionProcess
from scripts.percept import Percept
from scripts.learning_strategies.valueIteration import ValueIteration
mdp = MarkovDecisionProcess(0, range(16), range(4))
valueTest = ValueIteration(mdp, 0.8, 0.001, 0.90, 0.9, 1.0, 0.01)
percept = Percept(0, 2, 4, 2, False)
cProfile.run('for i in range(1000):' ' valueTest.evaluate(percept)')
def setUp(self):
self.mdp = MarkovDecisionProcess()
self.nstep = NstepQlearning(self.mdp, 0.8, 0.001, 0.7, 5, 1.0, 0.01)
self.nstep.mdp.reset()
self.nstep.evaluate(Percept(0, 2, 1, 1, False))
self.nstep.evaluate(Percept(0, 2, 4, 2, False))
def setUp(self):
self.mdp = MarkovDecisionProcess()
self.monte = MonteCarlo(self.mdp, 0.8, 0.001, 0.6, 1.0, 0.01)
self.monte.mdp.reset()
self.monte.evaluate(Percept(0, 2, 1, 1, False))
self.monte.evaluate(Percept(0, 2, 4, 2, False))
def setUp(self):
self.mdp = MarkovDecisionProcess()
self.mdp.reset()
self.mdp.update(Percept(0, 2, 1, 1, False))
self.mdp.update(Percept(0, 2, 4, 2, False))
def setUp(self):
self.mdp = MarkovDecisionProcess()
self.qlearning = Qlearning(self.mdp, 0.8, 0.01, 0.01, 1.0, 0.01)
self.qlearning.mdp.reset()
self.qlearning.evaluate(Percept(0, 2, 1, 1, False))
self.qlearning.evaluate(Percept(0, 2, 4, 2, False))