bgalbraith / bandits Goto Github PK

class EpsilonGreedyPolicy(Policy):
    [................................]

    def choose(self, agent):
        if np.random.random() < self.epsilon:
            return np.random.choice(len(agent.value_estimates))
        else:
            action = np.argmax(agent.value_estimates)                    <---------
            check = np.where(agent.value_estimates == action)[0]   <------
            if len(check) == 0:
                return action
            else:
                return np.random.choice(check)

def choose(self, agent):
        if np.random.random() < self.epsilon:
            return np.random.choice(len(agent.value_estimates))
        else:
            action = np.argmax(agent.value_estimates)                    <---------
            check = np.where(agent.value_estimates == agent.value_estimates[action])[0]   <------
            if len(check) == 1:   <--- At least there is going to be 1 
                return action
            else:    <---- Ties are solved randomly
                return np.random.choice(check)

class EpsilonGreedyPolicy(Policy):
    """
    The Epsilon-Greedy policy will choose a random action with probability
    epsilon and take the best apparent approach with probability 1-epsilon. If
    multiple actions are tied for best choice, then a random action from that
    subset is selected.
    """
    def __init__(self, epsilon):
        self.epsilon = epsilon

    def __str__(self):
        return '\u03B5-greedy (\u03B5={})'.format(self.epsilon)

    def choose(self, agent):
        if np.random.random() < self.epsilon:
            return np.random.choice(len(agent.value_estimates))
        else:
            action = np.argmax(agent.value_estimates)
            check = np.where(agent.value_estimates == action)[0]
            if len(check) == 0:
                return action
            else:
                return np.random.choice(check)