智能规划中面向简单偏好的高效求解方法

智能规划(AI planning)简称规划,是人工智能领域的一个重要分支,在各领域均有广泛应用,如工厂车间作业调度、物资运输调度、机器人动作规划以及航空航天任务规划等.传统智能规划要求规划解(动作序列)必须最终实现整个目标集合,这种目标一般被称为硬目标(hard goal).然而,许多实际问题中,求解的重点并不只是尽快实现目标以及尽量减少动作序列产生的代价,还需考虑其他因素,如资源消耗或时间约束等.为此,简单偏好(也称软目标soft goal)的概念应运而生.与硬目标相反,简单偏好是可以违背的.本质上,简单偏好用于衡量规划解质量的优劣,而不会影响规划解是否存在.现有关于简单偏好的研究进展缓慢,在规划解质量方面不尽如人意即求得的规划解与最优解的差距较大.提出了一种求解简单偏好的高效规划方法,将简单偏好表达为经典规划(classical planning)模型的一部分,并利用SMT (satisfiability modulo theories)求解器识别多个简单偏好之间的各种关系,从而约简简单偏好集,减轻规划器的求解负担.该方法的主要优势在于:一方面,提前对简单偏好集进行裁剪,在一定程度...     

模式数据库在智能规划中的应用

该规划器通过对智能规划领域里传统的构造模式数据库的方法进行改进,从而改进模式数据库启发式的效率：通过分析和移除一些在实际问题空间里不可能存在对应的完整状态的模式有效地减少了模式数据库的构造时间,并提高了模式数据库启发值的紧致性,使得模式数据库启发式能更好的指导搜索算法以求得问题的最优解。该规划器在linux系统下设计,通过使用规划器解决积木世界领域的规划问题来研究改进前后模式数据库启发式在搜索过程中所起的作用。     

基于不完全算法的并行FPGA SAT求解器

可满足性问题是计算机理论与应用的核心问题。在FPGA上提出了一个基于不完全算法的并行求解器pprobSAT+。使用多线程的策略来减少相关组件的等待时间,提高了求解器效率。此外,不同线程采用共用地址和子句信息的数据存储结构,以减少片上存储器的资源开销。当所有数据均存储在FPGA的片上存储器时,pprobSAT+求解器可以达到最佳性能。实验结果表明,相比于单线程的求解器,所提出的pprobSAT+求解器可获得超过2倍的加速比。     

一种近似模型表示的启发式Dyna优化算法

针对基于查询表的Dyna优化算法在大规模状态空间中收敛速度慢、环境模型难以表征以及对变化环境的学习滞后性等问题,提出一种新的基于近似模型表示的启发式Dyna优化算法(a heuristic Dyna optimization algorithm using approximate model representation, HDyna-AMR),其利用线性函数近似逼近Q值函数,采用梯度下降方法求解最优值函数.HDyna-AMR算法可以分为学习阶段和规划阶段.在学习阶段,利用agent与环境的交互样本近似表示环境模型并记录特征出现频率;在规划阶段,基于近似环境模型进行值函数的规划学习,并根据模型逼近过程中记录的特征出现频率设定额外奖赏.从理论的角度证明了HDyna-AMR的收敛性.将算法用于扩展的Boyan chain问题和Mountain car问题.实验结果表明,HDyna-AMR在离散状态空间和连续状态空间问题中能学习到最优策略,同时与Dyna-LAPS(Dyna-style planning with linear approximation and prioritized sweeping)和Sarsa(λ)相比,HDyna-AMR具有收敛速度快以及对变化环境的近似模型修正及时的优点.     

智能规划器StepByStep的研究和开发

智能规划器是智能规划研究成果的重要表现形式,规划器的求解效率和规划质量是智能规划理论研究的直接反映.首先介绍智能规划器的一般结构和StepByStep规划器的总体结构,然后详细阐述StepByStep规划器各组成部分所采用的方法和策略,定义谓词知识树来提取领域知识.在谓词知识树的基础上定义谓词规划树,并用各种策略来提高规划树的生成效率.在谓词规划树的基础上设计StepByStep的规划策略,最后用8个规划器对3个具有代表性的基准规划领域及其规划问题进行实际的求解实验,分析了StepByStep规划器在求解效率和规划质量上的具体表现.实验数据表明,StepByStep规划器的规划策略对3个不同规划领域都具有很好的指导作用,验证了领域知识在规划求解过程中的实际价值.     

基于模型检测的领域约束规划

基于模型检测的智能规划是当今通用的智能规划研究的热点,其求解效率比较高.但是,目前基于模型检测的智能规划系统没有考虑到利用领域知识来提高描述能力和求解效率.为此,研究了增加领域约束的基于模型检测的智能规划方法,并据此建立了基于模型检测的领域约束规划系统DCIPS(domain constraints integrated planning system).它主要考虑了领域知识在规划中的应用,将领域知识表示为领域约束添加到规划系统中.根据"规划=动作+状态",DCIPS将领域约束分为3种,即对象约束、过程约束和时序约束,采用对象约束来表达状态中对象之间的关系,采用过程约束来表达动作之间的关系,采用时序约束表达动作与状态中对象之间的关系.通过在2002年智能规划大赛AIPS 2002上关于交通运输领域的3个例子的测试,实验结果表明,利用领域约束的DCIPS可以方便地增加领域知识,更加实用化,其效率也有了相应的提高.     

结束时间不确定的投资组合选择问题建模与模型求解方法

针对结束时间具有不确定性的投资问题,建立以区间风险值(PVaR)度量市场风险的收益最大化投资组合选择模型.PVaR计算的复杂性使得模型难以运用一般优化方法求解,因此提出并证明可以通过求解等效的混合整数规划模型来得到原模型的最优解.利用实际股价数据进行数值实验分析,结果表明,求解混合整数规划模型针对小规模短期投资问题可以快速给出最优投资决策方案.     

基于流形学习的正交稀疏保留投影鉴别分析

稀疏保留投影( SPP)是一种保留样本间的稀疏重构关系的特征提取方法。但是根据流形学习理论,考虑局部流形结构比考虑全局欧氏结构更重要。此外,SPP得到的不是一组正交的投影向量,特征间存在冗余信息。为解决该问题,文中提出一种改进的稀疏保留投影算法,在SPP中引入有监督的流形学习,使得所得投影空间正交,并用迭代的方式求解最优投影变换,称为基于流形学习的迭代正交稀疏保留鉴别分析( MLIOSDA)。同时提出一种终止准则终止迭代。在CAS-PEAL人脸数据库和PolyU掌纹数据库的实验结果表明,文中提出的方法与一些相关方法相比有效地提高了识别结果。     

杂合启发式在线POMDP 规划

许多不确定环境下的自主机器人规划任务都可以用部分可观察的马氏决策过程(partially observableMarkov decision process,简称POMDP)建模.尽管研究者们在近似求解技术的设计方面已经取得了显著的进展,开发高效的POMDP 规划算法依然是一个具有挑战性的问题.以前的研究结果表明:在线规划方法能够高效地处理大规模的POMDP 问题,因而是一类具有研究前景的近似求解方法.这归因于它们采取的是“按需”作决策而不是预前对整个状态空间作决策的方式.旨在通过设计一个新颖的杂合启发式函数来进一步加速POMDP 在线规划过程,该函数能够充分利用现有算法里一些被忽略掉的启发式信息.实现了一个新的杂合启发式在线规划(hybrid heuristiconline planning,简称HHOP)算法.在一组POMDP 基准问题上,HHOP 有明显优于现有在线启发式搜索算法的实验性能.     

基于动态交通仿真模型的最优路径选择方法*

采用动态交通仿真模型INTEGRATION搭建了动态交通仿真平台,应用组件式蚁群算法来求解动态交通信息诱导下的最优路径选择问题。实例表明,基于动态交通仿真模型的最优路径选择方法是可行的、正确的和有效的。该方法易于理解和使用,具有很强的可重用性和可扩展性,为求解各类优化问题提供了可持续发展的框架。     

Using Many-Sorted Logic in the Object-Oriented Data Model for Fast Robot Task Planning

Search space explosion is a critical problem in robot task planning. This problem limits current robot task planners to solve only simple block world problems and task planning in a real robot working environment to be impractical. This problem is mainly due to the lack of utilization of domain information in task planning. In this paper, we describe a fast task planner for indoor robot applications that effectively uses domain information to speed up the planning process. In this planner, domain information is explicitly represented in an object-oriented data model (OODM) that uses many-sorted logic (MSL) representation. The OODM is convenient for the management of complex data and many-sorted logic is effective for pruning in the rule search process. An inference engine is designed to take advantage of the salient features of these two techniques for fast task planning. A simulation example and complexity analysis are given to demonstrate the advantage of the proposed task planner.     

An enhanced genetic algorithm for automated assembly planning

Automated assembly planning reduces manufacturing manpower requirements and helps simplify product assembly planning, by clearly defining input data, and input data format, needed to complete an assembly plan. In addition, automation provides the computational power needed to find optimal or near-optimal assembly plans, even for complex mechanical products. As a result, modern manufacturing systems use, to an ever greater extent, automated assembly planning rather than technician-scheduled assembly planning. Thus, many current research reports describe efforts to develop more efficient automated assembly planning algorithms. Genetic algorithms show particular promise for automated assembly planning. As a result, several recent research reports present assembly planners based upon traditional genetic algorithms. Although prior genetic assembly planners find improved assembly plans with some success, they also tend to converge prematurely at local-optimal solutions. Thus, we present an assembly planner, based upon an enhanced genetic algorithm, that demonstrates improved searching characteristics over an assembly planner based upon a traditional genetic algorithm. In particular, our planner finds optimal or near-optimal solutions more reliably and more quickly than an assembly planner that uses a traditional genetic algorithm.     

Learning Hierarchical Task Models from Input Traces

We describe HTN‐MAKER , an algorithm for learning hierarchical planning knowledge in the form of task‐reduction methods for hierarchical task networks (HTNs). HTN‐MAKER takes as input a set of planning states from a classical planning domain and plans that are applicable to those states, as well as a set of semantically annotated tasks to be accomplished. The algorithm analyzes this semantic information to determine which portion of the input plans accomplishes a particular task and constructs task‐reduction methods based on those analyses. We present theoretical results showing that HTN‐MAKER is sound and complete. Our experiments in five well‐known planning domains confirm the theoretical results and demonstrate convergence toward a set of HTN methods that can be used to solve any problem expressible as a classical planning problem in that domain, relative to a set of goal types for which tasks have been defined. In three of the five domains, HTN planning with the learned methods scales much better than a modern classical planner.     

基于学习的规划技术研究

经过近十多年的努力,现代智能规划器无论是效率还是处理能力均得到了极大提高。鉴于现有规划理论的局限性,进一步提高现有规划技术效率已愈显困难。现有的大多数规划器均不具备学习能力,无法从先前求解经验中学习有用知识。综述了基于学习的规划技术的发展现状,然后重点介绍了规划大赛中最佳学习器所使用的学习技术,最后指出当前基于学习的规划技术研究领域中存在的主要问题。     

Planning with Multistep Forward Search with Forced Goal‐Ordering Constraints

To solve a real‐world planning problem with interfering subgoals, it is essential to perform early detection of subgoal dependencies and achieve the subgoals in the correct order. This is also the case for planning problems with forced goal‐ordering (FGO) constraints. In automated planning, forward search with FGO constraints has been proposed many times over the years, but there are still major difficulties in realizing these FGOs in plan generation. Many existing methods such as goal agenda manager and ordered landmarks cannot detect the FGOs accurately, and thus, the undiscovered ordering relationship may cause the forward search to suffer from deadlocks. In this article, we put forward an approach via an effective search heuristic to constrain a planner to satisfy the FGOs. We make use of an atomic goal‐achievement graph in a look‐ahead search under the FGO constraints. This allows a forward search strategy to plan forward efficiently in multiple steps toward a goal state along a search path. Experimental results illustrate that, by avoiding deadlocks, we can solve more benchmark planning problems more efficiently than previous approaches. We also prove several formal properties for search that are related to FGO detection.     

Learning to solve planning problems efficiently by means of genetic programming.

Declarative problem solving, such as planning, poses interesting challenges for Genetic Programming (GP). There have been recent attempts to apply GP to planning that fit two approaches: (a) using GP to search in plan space or (b) to evolve a planner. In this article, we propose to evolve only the heuristics to make a particular planner more efficient. This approach is more feasible than (b) because it does not have to build a planner from scratch but can take advantage of already existing planning systems. It is also more efficient than (a) because once the heuristics have been evolved, they can be used to solve a whole class of different planning problems in a planning domain, instead of running GP for every new planning problem. Empirical results show that our approach (EvoCK) is able to evolve heuristics in two planning domains (the blocks world and the logistics domain) that improve PRODIGY4.0 performance. Additionally, we experiment with a new genetic operator --Instance-Based Crossover--that is able to use traces of the base planner as raw genetic material to be injected into the evolving population.