The creation and maintenance of a domain model is a well recognised bottleneck in the use of automated planning; indeed, ensuring a planning engine is fed with an accurate model of an application is essential in order that generated plans are effective. Engineering domain models using a hybrid representation is particularly challenging as it requires accurately describing continuous processes, which can have complex numeric effects. In this work we consider the problem of the refinement of an engineered hybrid domain model, to more accurately capture the effect of the underlying processes. Our approach exploits the information content of the original model, utilising machine learning techniques to identify important situation and temporal features that indicate a variation in the original effect. We use the problem of modelling traffic flows in an Urban Traffic Management setting as a case study and demonstrate in our evaluation that the refined domain models provide more accurate simulation, which can lead to higher quality plans.


Reprocess: Process Refinement for Improving Accuracy in Hybrid Planning Domain Models

This paper introduce a domain modeling method for multi-payload planning of experimental satellite based on temporal planning paradigm, which may express temporal constraints and corresponding handlers in a universal style. This will allow user to model planning problem more efficiently, and guide planning process throw domain description files.

Domain Modeling for Multi-Payload Planning of Experimental Satellite

Approaches based on deep network models are increasingly being used for pattern recognition and decision-making tasks in robotics and AI.  These approaches are characterized by a large labeled dataset, high computational complexity, and difficultly in understanding the internal representations and reasoning mechanisms. As a step towards addressing these limitations, our architecture uses non-monotonic logical reasoning with incomplete commonsense domain knowledge, and inductive learning of previously unknown state constraints, to guide the construction of deep networks based on a small number of training examples. As an illustrative example, we consider a robot reasoning about the stability and partial occlusion of object configurations in simulated images of an indoor domain. Experimental results indicate that in comparison with an architecture based just on deep networks, our architecture improves reliability, and reduces the sample complexity and time complexity of training the deep networks.

Commonsense Reasoning and Knowledge Acquisition to Guide Deep Learning on Robots

This paper focuses on the under-studied problem of automated decentralized AI planning under severe communications and information-sharing restrictions between the planners during planning. Under such conditions, we will argue that traditional logic-based knowledge representations in planning systems are not only brittle and hard to manipulate for reasoning and deliberations but also not suitable for conveying important and maximal information with minimal communications
between the planners or one's observations of another. We describe our proposed representation for object-centric knowledge graphs as a first-class element in our decentralized Hierarchical Task Network (HTN) based decentralized planning framework, called ARCADE. We present a number of preliminary experimental results, enabled by our new formalism.

Object-Centric Knowledge Representations in Hierarchical Planning for Decentralized Environments

The verification of planning domain models is crucial to ensure the safety, integrity and correctness of planning-based automated systems. This task is usually performed using model checking techniques.  However, directly applying model checkers to verify planning domain models can result in false positives, i.e. counterexamples that are unreachable by a sound planner when using the domain under verification during a planning task. In this paper, we discuss the downside of unconstrained planning domain model verification. We then propose a fail-safe practice for designing planning domain models that can inherently guarantee the safety of the produced plans in case of undetected errors in domain models.  In addition, we demonstrate how model checkers, as well as state trajectory constraints planning techniques, should be used to verify planning domain models so that unreachable counterexamples are not returned.

Goal-constrained planning domain model formal verification of safety properties

The International Competition on Knowledge Engineering for Planning and Scheduling (ICKEPS) plays a pivotal role in fostering the development of new Knowledge Engineering (KE) tools, and in emphasising the importance of principled approaches for all the different KE aspects that are needed for the successful long-term use of planning in real-world applications.  
In this paper, as an exercise in synthesis and for the sake of stimulating thoughts and discussion, we review the format of previous ICKEPS, to suggest alternative formats for future competitions, ideally to motivate someone to step up and organise the next ones.

The International Competition on Knowledge Engineering for Planning and Scheduling: Food for Thoughts (and Call to Action)

Learning for planning has extensive aspirations to aim at. Our primary focus area of review in this paper is the knowledge engineering of domain models to feed planners. This concerns with learning and representation of knowledge about operator schema, discrete or continuous resources, processes and events involved. With the increased problem complexity on the continuum of planning from classical towards full scope, synthesizing expressive and intensive planning domain models by hand become more challenging, time-intense and error-prone. Our intended contribution is to provide a broader perspective on the range of research in the domain model learning area and specification of interdependent AP dimensions which can affect the developmental decisions of learning methods/techniques. 
We map the amplitude of recently created domain model learning methods on five considerable dimensions. The dimensions to underpin the development of a learning system include the type of planning paradigm, representation languages, learning at various planning stages, learning systems and sources and the extent of learning that takes place.  
We also identify considerable issues for future work.

Considerable Factors for Domain Model Learning in Planning

The soundness and optimality of a plan depends on the correctness of the domain model. In real-world applications, specifying complete domain models is difficult as the interactions between the agent and its environment can be quite complex. We propose a framework to learn a PPDDL representation of the model incrementally over multiple planning problems using only experiences from the current planning problem, which suits non-stationary environments. We introduce the novel concept of reliability as an intrinsic motivation for reinforcement learning, and as a means of learning from failure to prevent repeated instances of similar failures. Our motivation is to improve both learning efficiency and goal-directedness. We evaluate our work with experimental results for three planning domains.

Incremental Learning of Action Models for Planning

Most approaches to learning action planning models heavily rely on a significantly large volume of training samples or plan observations. In this paper, we adopt a different approach based on deductive learning from domain-specific knowledge, specifically from logic formulae that specify constraints about the possible states of a given domain. The minimal input observability required by our approach is a single example composed of a full initial state and a partial goal state. We will show that exploiting specific domain knowledge enables to constrain the space of possible action models as well as to complete partial observations, both of which turn out helpful to learn good-quality action models.

One-shot learning: From domain knowledge to action models

Epistemic Planning can be used to achieve implicit coordination in cooperative multi-agent settings where knowledge and capabilities are distributed between the agents. In these scenarios, agents plan and act on their own without having to agree on a common plan or protocol beforehand. However, epistemic planning is undecidable in general. In this paper, we identify a decidable fragment of epistemic planning that allows for arbitrary initial state uncertainty and nondeterminism, but where actions can never increase the uncertainty of the agents. We show that in this fragment, planning with and without implicit coordination can be reduced to fully observable nondeterministic (FOND) planning and that it shares the same omputational complexity. We also provide a small case study, modeling the problem of multi-agent path finding with destination uncertainty in FOND, to show that our compilation approach can be successfully applied in practice.

Planning for Implicit Coordination using FOND

We propose a framework for learning discrete deterministic planning domains. In this framework, an agent learns the domain by observing the action effects through continuous features that describe the state of the environment after the execution of each action. Besides, the agent learns its perception function, i.e., a probabilistic mapping between state variables and sensor data represented as a vector of continuous random variables called perception variables. We define an algorithm that updates the planning domain and the perception function by (i) introducing new states, either by extending the possible values of state variables, or by weakening their constraints; (ii) adapts the perception function to fit the observed data (iii) adapts the transition function on the basis of the executed actions and the effects observed via the perception function. The framework is able to deal with exogenous events that happen in the environment.

ICAPS 2019 Workshop KEPS Submissions