CPO_point_gather.py 文件源码

def run_task(*_):
        trpo_stepsize = 0.01
        trpo_subsample_factor = 0.2

        env = PointGatherEnv(apple_reward=10,bomb_cost=1,n_apples=2, activity_range=6)

        policy = GaussianMLPPolicy(env.spec,
                    hidden_sizes=(64,32)
                 )

        baseline = GaussianMLPBaseline(
            env_spec=env.spec,
            regressor_args={
                    'hidden_sizes': (64,32),
                    'hidden_nonlinearity': NL.tanh,
                    'learn_std':False,
                    'step_size':trpo_stepsize,
                    'optimizer':ConjugateGradientOptimizer(subsample_factor=trpo_subsample_factor)
                    }
        )

        safety_baseline = GaussianMLPBaseline(
            env_spec=env.spec,
            regressor_args={
                    'hidden_sizes': (64,32),
                    'hidden_nonlinearity': NL.tanh,
                    'learn_std':False,
                    'step_size':trpo_stepsize,
                    'optimizer':ConjugateGradientOptimizer(subsample_factor=trpo_subsample_factor)
                    },
            target_key='safety_returns',
            )

        safety_constraint = GatherSafetyConstraint(max_value=0.1, baseline=safety_baseline)



        algo = CPO(
            env=env,
            policy=policy,
            baseline=baseline,
            safety_constraint=safety_constraint,
            safety_gae_lambda=1,
            batch_size=50000,
            max_path_length=15,
            n_itr=100,
            gae_lambda=0.95,
            discount=0.995,
            step_size=trpo_stepsize,
            optimizer_args={'subsample_factor':trpo_subsample_factor},
            #plot=True,
        )

        algo.train()