Example #1
0
def main():
    # All agents use  a tabular model will initial values of 0
    # Updates are done via TD learning with a fixed learning rate
    # Action_FA of discrete max means the agent chooses the action with the highest utility from a discrete array
    base_agent = Agent(model=TabularModel(mean=0, std=0),
                       action_fa=DiscreteMaxFA(),
                       optimiser=TemporalDifference(learning_rate=FixedDecay(0.2)))

    # Randomly select the next action
    random_agent = copy.deepcopy(base_agent)
    random_agent.exploration = RandomExploration()

    # Always select the best action seen so far (is default behaviour for agents)
    greedy_agent = copy.deepcopy(base_agent)

    # Always select the best action seen so far with optimistic starting values
    optimistic_greedy_agent = copy.deepcopy(base_agent)
    optimistic_greedy_agent.model = TabularModel(mean=1, std=0)

    # Select a random action with decaying likelihood
    egreedy_agent = copy.deepcopy(base_agent)
    egreedy_agent.exploration = EpsilonGreedy(FixedDecay(1, 0.995, 0.01))

    # Select a random action with fixed likelihood
    fixed_egreedy_agent = copy.deepcopy(base_agent)
    fixed_egreedy_agent.exploration = EpsilonGreedy(FixedDecay(0.2))

    # Explores using softmax
    boltzmann_agent = copy.deepcopy(base_agent)
    boltzmann_agent.exploration = Softmax(FixedDecay(2, 0.995, 0.1))

    agents = [random_agent, greedy_agent, optimistic_greedy_agent, egreedy_agent, fixed_egreedy_agent, boltzmann_agent]
    labels = ['Random', 'Greedy', 'Optimistic Greedy', 'E-Greedy Decay', 'E-Greedy Fixed', 'Boltzmann']

    agent_reward = []
    max_reward = []
    episodes = 100

    for agent in agents:
        path = "/tmp/rlagents/"
        am = AgentManager(agent=agent)
        em = EnvManager('BanditTenArmedUniformDistributedReward-v0', am)
        em.run(n_episodes=episodes, print_stats=False, path=path, video_callable=False)

        max_reward.append(max(em.env.r_dist))
        results = load_results(path)
        agent_reward.append(results['episode_rewards'])

    for i, ar in enumerate(agent_reward):
        percent_correct = [agent_reward[i][:j].count(max_reward[i])/float(j) for j in range(1, episodes)]
        plt.plot(range(1, episodes), percent_correct, label=labels[i])

    plt.xlabel('Steps')
    plt.ylabel('% Optimal Arm Pulls')
    plt.ylim(-0.2, 1.5)
    plt.legend(loc=2)

    plt.show()
Example #2
0
def test_env_reuse():
    with helpers.tempdir() as temp:
        env = gym.make('CartPole-v0')
        env.monitor.start(temp)
        env.monitor.close()

        env.monitor.start(temp, force=True)
        env.reset()
        env.step(env.action_space.sample())
        env.step(env.action_space.sample())
        env.monitor.close()

        results = monitor.load_results(temp)
        assert results['episode_lengths'] == [2], 'Results: {}'.format(results)