The work on heuristic planning concerns the development of heuristic mechanisms that can augment state space search (following the STRIPS paradigm) in order to speed up the planning process and produce better solutions (shorter plans). The research in this area has resulted in a number of pruning techniques, heuristic functions and algorithms that have been utilized by the following systems:
a) HAP: a domain-independent, state-space heuristic planning system, which can be customized through 7 planning parameters. HAP is a general planning platform which integrates the search modules of the BP planner, the heuristics of AcE and several techniques for speeding up the planning process. Apart from the selection of the planning direction, which is the most important feature of HAP, the user can also set the values of 6 other parameters that mainly affect the search strategy and the heuristic function.
b) AcE: a domain-independent, state-space heuristic planning system, which obtains estimates for the cost of applying each action of the domain by performing a forward search in a relaxed version of the initial problem in order to come up with a heuristic function that is utilized in a backward search on the original problem. The system is further enhanced by a goal-ordering technique.
c) GRT: a heuristic state-space planner that constructs its heuristic function in a domain-independent way. The planner achieves significant performance in many domains as it has been shown in international planning competitions. An extension to the basic planning system uses XOR-constraints, in order to analyze a planning problem in a sequence of easier sub-problems that have to be solved sequentially. XOR-constraints are relations between sets of ground facts, where exactly one of them can hold in each state.
The main research in parallel planning has been focused on ways to parallelize the planning process of any strips planner without altering the resulting plans. We have developed a method called ODMP (Operator Distribution Method for parallel Planning) based on the distribution of semi-grounded operators to the available processors, which manages to speedup any heuristic strips planner while preserving the quality of the resulting plans. The method has been tested on GRT and CL, a domain-dependent strips planner for logistics problems, with remarkable success.
Planning for Web Service Composition
The research in this area focuses on ways to utilize AI planning techniques in order to automatically create composite web services from atomic ones. The main outcome is the PORSCE system, which uses Planning, Ontologies and Reasoning techniques in order to compose atomic web services according to the user defined inpouts and outputs. The system provides the user with a graphical user interface for creating, visualizing and editting composite services.
Machine Learning aided Planning
Research in this area is concerned with the enhancement of planning systems using techniques from Machine Learning in order to automatically configure their planning parameters according to the morphology of the problem in hand. More specifically two different adaptive systems that set the planning parameters of a highly adjustable planner based on measurable characteristics of the problem instance have been developed. These systems have acquired their knowledge from a large data set produced by results from experiments on many problems from various domains:
a) HAPrc: The first planner is a rule-based system that employs propositional rule learning to induce knowledge that suggests effective configuration of planning parameters based on the problem’s characteristics.
b) HAPnn: The second planner employs instance-based learning in order to find problems with similar structure and adopt the planner configuration that has proved in the past to be effective on these problems.
Planning for Curricula Synthesis
The research in this area focuses on ways to utilize AI techniques in order to create automatic or semi-automatic curricula synthesis systems. The main outcome of the research is the PASERsystem, which uses AI planning and Ontologies in order to dynamically compose learning objects according to the learner’s knowledge state, educational goals, preferences and abilities, together with the available learning material. PASER translates the synthesis problem as a planning problem, where the learner’s knowledge state forms the Initial state, the educational goals form the goals of the problem and the available educational material form the problem actions. The connections among these elemetns are made feasible through the use of an ontology describing educational terms and relations.