one_step_ac.py 文件源码

def run(self, 
            epochs,  
            steps, 
            api_key,
            rollouts_per_epoch = 20,
            updateTargetNetwork = defaultRunSettings['updateTargetNetwork'], 
            explorationRate = defaultRunSettings['explorationRate'], 
            miniBatchSize = defaultRunSettings['miniBatchSize'], 
            learnStart = defaultRunSettings['learnStart'], 
            renderPerXEpochs = defaultRunSettings['renderPerXEpochs'], 
            shouldRender = defaultRunSettings['shouldRender'], 
            experimentId = defaultRunSettings['experimentId'], 
            force = defaultRunSettings['force'], 
            upload = defaultRunSettings['upload']):

        last100Scores = [0] * 100
        last100ScoresIndex = 0
        last100Filled = False

        stepCounter = 0

        if not experimentId == None:
            self.env.monitor.start('tmp/'+experimentId, force = force)

        for epoch in xrange(epochs):
            I = 1
            observation = self.env.reset();
            for t in xrange(steps):
                policyValues = self.runModel(self.policyModel, observation)
                action = self.selectActionByProbability(policyValues)

                newObservation, reward, done, info = self.env.step(action)

                cost, grads = self.get_cost_grads(self.policyModel);
                print (theano.pp(grads[1][0]));

                if done:
                    delta = reward + self.discountFactor * self.runModel(self.valueModel, newObservation) - self.runModel(self.valueModel, observation)
                else :
                    delta = reward - self.runModel(self.valueModel, observation) # because the value for new obs is 0

        self.env.monitor.close()
        if upload:
            gym.upload('/tmp/'+experimentId, api_key=api_key)