Blazing the trails before beating the path: Sample-efficient Monte-Carlo planning