当前位置: 首页>>代码示例>>Python>>正文


Python special.discount_cumsum方法代码示例

本文整理汇总了Python中rllab.misc.special.discount_cumsum方法的典型用法代码示例。如果您正苦于以下问题:Python special.discount_cumsum方法的具体用法?Python special.discount_cumsum怎么用?Python special.discount_cumsum使用的例子?那么, 这里精选的方法代码示例或许可以为您提供帮助。您也可以进一步了解该方法所在rllab.misc.special的用法示例。


在下文中一共展示了special.discount_cumsum方法的4个代码示例,这些例子默认根据受欢迎程度排序。您可以为喜欢或者感觉有用的代码点赞,您的评价将有助于系统推荐出更棒的Python代码示例。

示例1: _worker_rollout_policy

# 需要导入模块: from rllab.misc import special [as 别名]
# 或者: from rllab.misc.special import discount_cumsum [as 别名]
def _worker_rollout_policy(G, args):
    sample_std = args["sample_std"].flatten()
    cur_mean = args["cur_mean"].flatten()
    K = len(cur_mean)
    params = np.random.standard_normal(K) * sample_std + cur_mean
    G.policy.set_param_values(params)
    path = rollout(G.env, G.policy, args["max_path_length"])
    path["returns"] = discount_cumsum(path["rewards"], args["discount"])
    path["undiscounted_return"] = sum(path["rewards"])
    if args["criterion"] == "samples":
        inc = len(path["rewards"])
    elif args["criterion"] == "paths":
        inc = 1
    else:
        raise NotImplementedError
    return (params, path), inc 
开发者ID:bstadie,项目名称:third_person_im,代码行数:18,代码来源:cem.py

示例2: sample_return

# 需要导入模块: from rllab.misc import special [as 别名]
# 或者: from rllab.misc.special import discount_cumsum [as 别名]
def sample_return(G, params, max_path_length, discount):
    # env, policy, params, max_path_length, discount = args
    # of course we make the strong assumption that there is no race condition
    G.policy.set_param_values(params)
    path = rollout(
        G.env,
        G.policy,
        max_path_length,
    )
    path["returns"] = discount_cumsum(path["rewards"], discount)
    path["undiscounted_return"] = sum(path["rewards"])
    return path 
开发者ID:bstadie,项目名称:third_person_im,代码行数:14,代码来源:cma_es.py

示例3: _worker_rollout_policy

# 需要导入模块: from rllab.misc import special [as 别名]
# 或者: from rllab.misc.special import discount_cumsum [as 别名]
def _worker_rollout_policy(G, args):
    sample_std = args["sample_std"].flatten()
    cur_mean = args["cur_mean"].flatten()
    n_evals = args["n_evals"]
    K = len(cur_mean)
    params = np.random.standard_normal(K) * sample_std + cur_mean
    G.policy.set_param_values(params)
    paths, returns, undiscounted_returns = [], [], []
    for _ in range(n_evals):
        path = rollout(G.env, G.policy, args["max_path_length"])
        path["returns"] = discount_cumsum(path["rewards"], args["discount"])
        path["undiscounted_return"] = sum(path["rewards"])
        paths.append(path)
        returns.append(path["returns"])
        undiscounted_returns.append(path["undiscounted_return"])
    
    result_path = {'full_paths':paths}
    result_path['undiscounted_return'] = _get_stderr_lb(undiscounted_returns)
    result_path['returns'] = _get_stderr_lb_varyinglens(returns)
       
    # not letting n_evals count towards below cases since n_evals is multiple eval for single paramset
    if args["criterion"] == "samples":
        inc = len(path["rewards"])
    elif args["criterion"] == "paths":
        inc = 1
    else:
        raise NotImplementedError
    return (params, result_path), inc 
开发者ID:shaneshixiang,项目名称:rllabplusplus,代码行数:30,代码来源:cem.py

示例4: compute_and_apply_importance_weights

# 需要导入模块: from rllab.misc import special [as 别名]
# 或者: from rllab.misc.special import discount_cumsum [as 别名]
def compute_and_apply_importance_weights(self,path):
        new_logli = self.algo.policy.distribution.log_likelihood(path["actions"],path["agent_infos"])
        logli_diff = new_logli - path["log_likelihood"]
        if self.algo.decision_weight_mode=='pd':
            logli_diff = logli_diff[::-1]
            log_decision_weighted_IS_coeffs = special.discount_cumsum(logli_diff,1)
            IS_coeff = np.exp(log_decision_weighted_IS_coeffs[::-1])
        elif self.algo.decision_weight_mode=='pt':
            IS_coeff = np.exp(np.sum(logli_diff))
        if self.algo.clip_IS_coeff_above:
            IS_coeff = np.minimum(IS_coeff,self.algo.IS_coeff_upper_bound)
        if self.algo.clip_IS_coeff_below:
            IS_coeff = np.maximum(IS_coeff,self.algo.IS_coeff_lower_bound)

        path["IS_coeff"] = IS_coeff 
开发者ID:jachiam,项目名称:cpo,代码行数:17,代码来源:sampler_safe.py


注:本文中的rllab.misc.special.discount_cumsum方法示例由纯净天空整理自Github/MSDocs等开源代码及文档管理平台,相关代码片段筛选自各路编程大神贡献的开源项目,源码版权归原作者所有,传播和使用请参考对应项目的License;未经允许,请勿转载。