Search

Intelligent agents must have a model of the dynamics of the domain in which they act. Models can be encoded by human experts or, as required by autonomous systems, automatically acquired from observation. At the state of the art, there existseveral systems for automated acquisition of planning domain models.\ud \ud \ud In this paper we present a brief overview of the automated tools that can be exploited to induce planning domain models. While reviewing the literature on the existing tools for Knowledge Engineering (KE), we do a comparative analysis of them. The analysis is based on a set of criteria.\ud \ud \ud The aim of the analysis is to give insights into the strengths and weaknesses of the considered systems, and to provide input for new, forthcoming research on KE tools in order to address future challenges in the automated KE area.

The ParLPG planning system is based on the idea of using a generic algorithm configuration procedure – here, the well-known ParamILS framework – to optimise the performance of a highly parametric planner on a set of problem instances representative of a specific planning domain.\ud This idea is applied to LPG, a versatile and efficient planner based on stochastic local-search with 62 parameters and over 6.5 × 10^17 possible configurations. A recent, large-scale empirical investigation showed that the approach behind ParLPG yields substantial performance improvements across a broad range of planning domains.

When designing state-of-the-art, domain-independent planning systems, many decisions have to be made with respect to the domain analysis or compilation performed during preprocessing, the heuristic functions used during search, and other\ud features of the search algorithm. These design decisions can have a large impact on the performance of the resulting planner. By providing many alternatives for these choices and exposing them as parameters, planning systems can in principle be configured to work well on different domains. However, usually planners are used in default configurations that have been chosen because of their good average performance over a set of benchmark domains, with limited experimentation of the potentially huge range of possible configurations.\ud In this work, we propose a general framework for automatically configuring a parameterised planner, showing that substantial performance gains can be achieved. We apply the framework to the well-known LPG planner, which has 62 parameters and over 6.5 × 10^17 possible configurations. We demonstrate that by using this highly parameterised planning system in combination with the off-the-shelf, state-of-the-art automatic algorithm configuration procedure ParamILS, substantial performance improvements on specific planning domains can be obtained.

We present PbP2, an automated system that generates efficient domain-specific multi planners from a portfolio of domain-independent planning techniques by (i) computing some sets of macro-actions for every planner in the portfolio, (ii) optimizing the parameter setting of the parameterized planners in the portfolio, (iii) selecting a promising combination of planners in the portfolio and relative useful macro-actions, and (iv) defining some running time slots for their round-robin scheduling during planning. The configuration of the portfolio yielding the multi planner relies on some knowledge about the performance of the planners and relative macro actions, which is automatically generated from a training problem set. PbP2 is a revision and extension of a preliminary version of this system (PbP) that was awarded at the learning track of IPC-2008.

Planning based on propositional satisfiability is a powerful approach for computing makespan-optimal plans. However, it is usually slower then heuristic-based sub-optimal approaches. In this work we propose MacroSatPlan; a SatPlan based planner which exploits macros extracted by Macro-FF and uses a predictive model of the optimal solution length that is constructed by WEKA, a commonly used toolkit of machine learning algorithms.

The paper raises some issues relating to the engineering of domain models for automated planning. It studies the idea of a domain model as a formal specification of a domain, and considers properties of that specification. It proposes some definitions, which the planning and, more generally, the artificial intelligence community needs to consider, in order to properly deal with engineering issues in domain model creation.

Local authorities that manage traffic-related issues in\ud urban areas have to optimise the use of available resources,\ud in order to minimise congestion and delays.\ud In this context, Automated Planning and Scheduling\ud can be fruitfully exploited, in order to provide dynamic\ud plans that help managing the urban road network.\ud In this paper we provide a review of existing planning\ud and scheduling approaches that have been designed for\ud dealing with different aspects of traffic management,\ud with the aim of gaining insights on the limits of current\ud applications, and highlighting the open challenges.

Domain-independent planning systems require that domain\ud constraints and invariants are specified as part of\ud the input domain model. In AI Planning, the generated\ud plan is correct provided the constraints of the world\ud in which the agent is operating are satisfied. Specifying\ud operator descriptions by hand for planning domain\ud models that also require domain specific constraints is\ud time consuming, error prone and still a challenge for the\ud AI planning community.\ud The LOCM (Cresswell, McCluskey, and West 2013)\ud system carries out automated generation of the dynamic\ud aspects of a planning domain model from a set of example\ud training plans. We enhance the output domain\ud model of the LOCM system to capture static domain\ud constraints from the same set of input training plans as\ud used by LOCM to learn dynamic aspects of the world.\ud In this paper we propose a new framework ASCoL (Automated\ud Static Constraint Learner), to make constraint\ud acquisition more efficient, by observing a set of training\ud plan traces. Most systems that learn constraints automatically\ud do so by analysing the operators of the planning\ud world. Out proposed system will discover static\ud constraints by analysing plan traces for correlations in\ud the data. To do this an algorithm is in the process of\ud development for graph discovery from the collection of\ud ground action instances used in the input plan traces.\ud The proposed algorithm will analyse the complete set of\ud plan traces, based on a predefined set of constraints, and\ud deduces facts from it. We then augment components of\ud the LOCM generated domain with enriched constraints.