class EpsilonGreedySDExplorer(DecayExplorer):
# define the conditions of the environment
inConditions = {'discreteStates': False, 'discreteActions':True}
# define the conditions of the environment
outConditions = {}
def setExperiment(self, experiment):
DecayExplorer.setExperiment(self, experiment)
# create random exploration mapping
self.explMapping = Linear(self.experiment.conditions['stateDim'], self.experiment.conditions['actionNum'])
self.randomizeMapping()
def applyState(self, state):
""" save current state for _explore() method later on. """
DecayExplorer.applyState(self, state)
self.state = state
return state
def _explore(self, action):
""" add an episode-specific offset to each action """
if np.random.random() < self.epsilon:
exploration = self.explMapping.predict(self.state).flatten()
action = np.argmax(exploration)
return action
def randomizeMapping(self):
self.explMapping.parameters = np.random.normal(0., 0.1, size=self.explMapping.parameters.shape)
def applyEpisodeFinished(self, episodeFinished):
""" apply transformations to episodeFinished and return it. """
DecayExplorer.applyEpisodeFinished(self, episodeFinished)
# at end of episode, randomize the exploration parameters
if episodeFinished:
self.randomizeMapping()
return episodeFinished
def setExperiment(self, experiment):
DecayExplorer.setExperiment(self, experiment)
# create random exploration mapping
self.explMapping = Linear(self.experiment.conditions['stateDim'], self.experiment.conditions['actionNum'])
self.randomizeMapping()
