Knowledge Engineering Tools and Techniques for AI Planning

The paper presents a classical planning compilation for learning STRIPS action models from partial observations of plan executions. The compilation is flexible to different amounts and types of input knowledge, from learning samples that comprise partially observed intermediate states of the plan execution to samples in which only the initial and final states are observed. The compilation accepts also partially specified action models and it can be used to validate whether an observation of a plan execution follows a given STRIPS action model, even if the given model or the given observation is incomplete.

Intelligent agents solving problems in the real world require domain models containing widespread knowledge of the world. Domain models can be encoded by human experts or automatically learned through the observation of some existing plans (behaviours). Encoding a domain model manually from experience and intuition is a very complex and time-consuming task, even for domain experts. This chapter investigates various classical and state-of-the-art methods proposed by the researchers to attain the ability of automatic learning of domain models from training data. This concerns with the learning and representation of knowledge about the operator schema, discrete or continuous resources, processes and events involved in the planning domain model. The taxonomy and order of these methods we followed are based on their standing and frequency of usage in the past research. Our intended contribution in this chapter is to provide a broader perspective on the range of techniques in the domain model learning area which underpin the developmental decisions of the learning tools.

The interest and scope of the area of autonomous systems have been steadily growing in the last 20 years. Artificial intelligence planning and scheduling is a promising technology for enabling intelligent behavior in complex autonomous systems. To use planning technology, however, one has to create a knowledge base from which the input to the planner will be derived. This process requires advanced knowledge engineering tools, dedicated to the acquisition and formulation of the knowledge base, and its respective integration with planning algorithms that reason about the world to plan intelligently. In this chapter, we shortly review the existing knowledge engineering tools and methods that support the design of the problem and domain knowledge for AI planning and scheduling applications (AI P&S). We examine the state-of-the-art tools and methods of knowledge engineering for planning & scheduling (KEPS) in the context of an abstract design process for acquiring, formulating, and analyzing domain knowledge. Planning quality is associated with requirements knowledge (pre-design) which should match properties of plans (post-design). While examining the literature, we analyze the design phases that have not received much attention, and propose new approaches to that, based on theoretical analysis and also in practical experience in the implementation of the system itSIMPLE.

The Planning Domain Definition Language (PDDL) is the state-of-the-art language for specifying planning problems in artificial intelligence research. Writing and maintaining these planning problems, however, can be time-consuming and error- prone. To address this issue, we present myPDDL—a modular toolkit for developing and manipulating PDDL domains and problems. To evaluate myPDDL, we compare its features to existing knowledge engineering tools for PDDL. In a user test, we additionally assess two of its modules, namely the syntax highlighting feature and the type diagram generator. The users of syntax highlighting detected 36% more errors than non-users in an erroneous domain file. The average time on task for questions on a PDDL type hierarchy was reduced by 48% when making the type diagram generator available. This implies that myPDDL can support knowledge engineers well in the PDDL design and analysis process.

In this chapter we describe the main pillars of the Planning.Domains initiative (API, Solver, Editor, and Education), detail some of the current use-cases for them, and outline the future path of the initiative. We further dive into some of the most recent developments of Planning.Domains, and shed light on what is next for the platform.

Domain-independent planning decouples planning task description, specified in a description language (e.g., PDDL), and planning engines that accept the task description as an input and generate plans (if they exist). A planning domain model gives general description of the environment and actions of a given domain while a planning problem specifies concrete objects, an initial state, and a goal. Planning domain model together with planning problem description forms a planning task. Hence it is typical that one domain model can be used for a class of planning tasks.

In mixed-initiative planning systems, humans and AI planners work together for generating satisfactory solution plans or making easier solving hard planning problems, which otherwise would require much greater human planning efforts or much more computational resources. In this approach to plan generation, it is important to have effective plan visualization capabilities, as well to support the user with some interactive capabilities for the human intervention in the planning process. This paper presents an implemented interactive tool for the visualization, generation, and revision of plans. The tool provides an environment through which the user can interact with a state-of-the-art domain-independent planner, and obtain an effective visualization of a rich variety of information during planning, including the reasons why an action is being planned or why its execution in the current plan is expected to fail, the trend of the resource consumption in the plan, and the temporal scheduling of the planned actions. Moreover, the proposed tool supports some ways of human intervention during the planning process to guide the planner towards a solution plan, or to modify the plan under construction and the problem goals.

Planning and scheduling are two closely related areas that deal with organizing activities to achieve a particular goal (planning) and allocating these activities to limited time and resources for execution (scheduling). However, regarding the tools supporting the planning and scheduling processes, these two areas are still far from each other. Progress in scheduling has been driven by industry and many techniques and tools to support the scheduling process have been designed. On the other hand, planning is still more an academic topic and, until recently, engineering support of the planning process has been limited. The focus of planning community was mainly on design of efficient planners, but this started to change in recent years and several tools supporting the planning process have been designed. This chapter focuses on interactive visualization of plans and schedules, that is, on the way how plans and schedules can be presented visually to users, and on tools that can work with these visualizations.

We describe a tool for providing explanation of plans to non-technical users, built on formal argumentation and dialogue theory, and supported by natural language generation and visualisation technologies. We describe how arguments can be generated from domain rules, and how justified arguments can be identified through dialogue, allowing the system to use such a dialogue to explain a plan. We provide information about our prototype system implementation, discussing its current limitations, and identifying potential avenues for future research.

We present LTS+ +, an interactive development environment for planning-based hypothesis generation motivated by applications that require multiple hypotheses to be generated in order to reason about the observations. Our system uses expert knowledge and AI planning to reason about possibly incomplete, noisy, or inconsistent observations derived from data by a set of analytics, and generates plausible and consistent hypotheses about the state of the world. Planning-based reasoning is enabled by knowledge models obtained from domain experts that describe entities in the world, their states, and relationship to observations. To address the knowledge engineering challenge, we have developed a language, also called LTS+ + that allows the domain expert to specify the state transition model and encoding of the observations without any knowledge of AI planning or existing planning languages (i.e., PDDL). LTS+ + integrated development environment facilitates model testing and debugging, generating, and visualizing multiple hypotheses for user-provided observations, and supports model deployment for online observation processing, publishing generated hypotheses for analysis by experts or other systems. To compute hypotheses we use an efficient planner that finds a set of high-quality plans. We experimentally evaluate our planning algorithm and conduct empirical evaluation to demonstrate the feasibility of our approach and the benefits of using planning-based reasoning. In this chapter we focus on describing the modeling and the knowledge engineering challenges of our system.

Automated planning tools are complex pieces of software that take declarative domain descriptions and generate plans from domains and problems. New users often find it challenging to understand the plan generation process, while experienced users often find it difficult to track semantic errors and efficiency issues. In response, we develop a cloud-based planning tool with code editing and state-space visualization capabilities that simplifies this process. The code editor focuses on visualizing the domain, problem, and resulting sample plan, helping the user see how such descriptions are connected without changing context. The visualization tool explores two alternative visualizations aimed at illustrating the operation of the planning process and how the domain dynamics evolve during plan execution.

During the last decade, industrial collaborative robots have entered assembly cells supporting human workers in repetitive and physical demanding operations. Such human-robot collaboration (HRC) scenarios entail many open issues. The deployment of highly flexible and adaptive plan-based controllers is capable of preserving productivity while enforcing human safety is then a crucial requirement. The deployment of plan-based solutions entails knowledge engineers and roboticists interactions in order to design well-suited models of robotic cells considering both operational and safety requirements. So, the ability of supporting knowledge engineering for integrating high level and low level control (also from non-specialist users) can facilitate deployment of effective and safe solutions in different industrial settings. In this chapter, we will provide an overview of some recent results concerning the development of a task planning and execution technology and its integration with a state of the art Knowledge Engineering environment to deploy safe and effective solutions in realistic manufacturing HRC scenarios. We will briefly present and discuss a HRC use case to demonstrate the effectiveness of such integration discussing its advantages.

Automated planning deals with the problem of finding a (partially ordered) action sequence, a plan, transforming the environment from a given initial state to some required goal state. In a nutshell, automated planning is a tool for deliberative reasoning which intelligent entities can use to find strategies (plans) for achieving longer-term goals. There are many successful real-world applications ranging from space and planet observations, Urban Traffic Control to narrative generation.

The core task of automated planning is goal-directed action selection; this task is not unique to the planning community, but is also relevant to numerous other research areas within AI. One such area is interactive systems, where a fundamental component called the interaction manager selects actions in the context of conversing with humans using natural language. Although this has obvious parallels to automated planning, using a planner to address the interaction management task relies on appropriate engineering of the underlying planning domain and planning problem to capture the necessary dynamics of the world, the agents involved, their actions, and their knowledge. In this chapter, we describe work on using domain-independent automated planning for action section in social human–robot interaction, focusing on work from the JAMES (Joint Action for Multimodal Embodied Social Systems) robot bartender project.