The most successful recent approaches to partially observable Markov decision problem (POMDP) solving have largely been point-based approximation algorithms. These work by selecting a finite number of belief points, computing alpha-vectors for those points, and using the resulting policy everywhere. However, if during execution the belief state is far from the points, there is no guarantee that the policy will be good. This case occurs either when the points are chosen poorly or there are too few points to capture the whole optimal policy, for example in domains where there are many low probability transitions, such as faults or exogenous events. In this paper we explore the use of an on-line plan repair approach to overcome this difficulty. The idea is to split computation between off-line plan creation and, if necessary, on-line plan repair. We evaluate a variety of heuristics used to determine when plan repair might be useful, and then repair the plan by sampling a small number of additional belief points and recomputing the policy. We show in several domains that the approach is more effective than either off-line planning alone even with much more computation time, or a purely on-line planning based on forward search. We also show that the overhead of checking the heuristics is very small when replanning is unnecessary.

In this paper, we revisit the idea of splitting a planning problem into subproblems hopefully easier to solve with the help of landmark analysis. While this technique initially proposed in the first approaches related to landmarks has been outperformed by landmark-based heuristics, we believe that it is still a promising research direction. To this end, we propose a new method for problem splitting based on landmarks which has two advantages over the original technique: it is complete (if a solution exists, the algorithm finds it), and it uses the precedence relation over the landmarks in a more flexible way. We lay in this paper the foundations of a meta best-first search algorithm, which explores the landmark orderings to create subproblems and can use any embedded planner to solve subproblems. It opens up avenues for future research: among them are new heuristics for guiding the meta search towards the most promising orderings, different policies for generating subproblems, and influence of the embedded subplanner.

AI planners have to compromise between the speed of the planning process and the quality of the generated plan. Anytime planners try to balance these objectives by finding plans of better quality over time, but current anytime planners often do not make effective use of increasing runtime beyond a certain limit. We present a new method of continuing plan improvement, that works by repeatedly decomposing a given plan into subplans and optimising each subplan locally. The decomposition exploits block-structured plan deordering to identify coherent subplans, which make sense to treat as units. This approach extends the "anytime capability" of current planners - to provide continuing plan quality improvement at any time scale.

One drawback of Hierarchical Task Network (HTN) planning is the difficulty of providing complete domain knowledge, i.e., a complete and correct set of HTN methods for every task. To provide a principled way to overcome this difficulty, we define a simple formalism that extends classical planning to include problem decomposition using methods, and a planning algorithm based on this formalism. In our formalism, the methods specify ways to achieve goals (rather than tasks as in conventional HTN planning), and goals may be achieved even when no methods are available. Our planning algorithm, GoDeL (Goal Decomposition with Landmarks), is sound and complete irrespective of whether the domain knowledge (i.e., the set of methods given to the planner) is complete. By comparing GoDeL's performance with varying amounts of domain knowledge across three benchmark planning domains, we show experimentally that (1) GoDeL works correctly with partial planning knowledge, (2) GoDeL's performance improves as more planning knowledge is given, and (3) when given full domain knowledge, GoDeL matches the performance of a state-of-the-art hierarchical planner.

Formulating knowledge for use in AI Planning engines is currently something of an ad-hoc process, where the skills of knowledge engineers and the tools they use may significantly influence the quality of the resulting planning application. There is little in the way of guidelines or standard procedures, however, for knowledge engineers to use when formulating knowledge into planning domain languages such as PDDL. This paper seeks to investigate this process using as a case study a road traffic accident management domain. Managing road accidents requires systematic, sound planning and coordination of resources to improve outcomes for accident victims. We have derived a set of requirements in consultation with stakeholders for the resource coordination part of managing accidents. We evaluate two separate knowledge engineering strategies for encoding the resulting planning domain from the set of requirements: (a) the traditional method of PDDL experts and text editor, and (b) a leading planning GUI with built in UML modelling tools. These strategies are evaluated using process and product metrics, where the domain model (the product) was tested extensively with a range of planning engines. The results give insights into the strengths and weaknesses of the approaches, highlight lessons learned regarding knowledge encoding, and point to important lines of research for knowledge engineering for planning.

The iPDB procedure by Haslum et al. is the state-of-the-art method for computing additive abstraction heuristics for domain-independent planning. It performs a hill-climbing search in the space of pattern collections, combining information from multiple patterns in the so-called canonical heuristic. We show how stronger heuristic estimates can be obtained through linear programming. An experimental evaluation demonstrates the strength of the new technique on the IPC benchmark suite.

A fault represents some erroneous operation of a system that could result from an action selection error or some abnormal condition. We formally define error models that characterize the likelihood of various faults and consider the problem of fault-tolerant planning, which optimizes performance given an error model. We show that factoring the possibility of errors significantly degrades the performance of stochastic planning algorithms such as LAO*, because the number of reachable states grows dramatically. We introduce an approach to plan for a bounded number of faults and analyze its theoretical properties. When combined with a continual planning paradigm, the k-fault-tolerant planning method can produce near-optimal performance, even when the number of faults exceeds the bound. Empirical results in two challenging domains confirm the effectiveness of the approach in handling different types of runtime errors.

Random walks have become a popular component of recent planning systems. The increased exploration is a valuable addition to more exploitative search methods such as Greedy Best First Search (GBFS). A number of successful planners which incorporate random walks have been built. The work presented here aims to exploit the experience gained from building those systems. It begins a systematic study of the design space and alternative choices for building such a system, and develops a new random walk planner from scratch, with careful experiments along the way. Four major insights are: 1. a high state evaluation frequency is usually superior to the endpoint-only evaluation used in earlier systems, 2. adjusting the restarting parameter according to the progress speed in the search space performs better than any fixed setting, 3. biasing the action selection towards preferred operators of only the current state is better than Monte Carlo Helpful Actions, which depend on the number of times an action has been a preferred operator in previous walks, and 4. even simple forms of random walk planning can compete with GBFS.

We propose a unified approach to plan execution and schedule dispatching that converts a plan, which has been augmented with temporal constraints, into a policy for dispatching. Our approach generalizes the original plan and temporal constraints so that the executor need only consider the subset of state that is relevant to successful execution of valid plan fragments. We can accommodate a variety of calamitous and serendipitous changes to the state of the world by supporting the seamless re-execution or omission of plan fragments, without the need for costly replanning. Our methodology for plan generalization and online dispatching is a novel combination of plan execution and schedule dispatching techniques. We demonstrate the effectiveness of our method through a prototype implementation and a series of experiments.

The ability to understand the goals and plans of other agents is an important characteristic of intelligent behaviours in many contexts. One of the approaches used to endow agents with this capability is the weighted model counting approach. Given a plan library and a sequence of observations, this approach exhaustively enumerates plan execution models that are consistent with the observed behaviour. The probability that the agent might be pursuing a particular goal is then computed as a proportion of plan execution models satisfying the goal. The approach allows to recognize multiple interleaved plans, but suffers from a combinatorial explosion of plan execution models, which impedes its application to real-world domains. This paper presents a heuristic weighted model counting algorithm that limits the number of generated plan execution models in order to recognize goals quickly by computing their lower and upper bound likelihoods.