Learners may develop expertise by experiencing numerous different but relevant situations. Computer games and virtual simulations can facilitate these training opportunities, however, because of the relative difficulty in authoring new scenarios, the increasing need for new and different scenarios becomes a bottleneck in the learning process. Furthermore, a one-size-fits-all scenario may not address all of the abilities, needs, or goals of a particular learner. To address these issues we present a novel technique, Automated Scenario Adaptation, to automatically “rewrite” narrative scenario content to suit individual learners’ needs and abilities and to incorporate recent changes from real world learning needs. Scenario adaptation acts as problem generation for intelligent tutoring systems, producing greater learning opportunities that facilitate engagement and continued learner involvement.

If given manually-crafted goal selection knowledge, goal reasoning agents can dynamically determine which goals they should achieve in complex environments. These agents should instead learn goal selection knowledge through expert interaction. We describe T-ARTUE, a goal reasoning agent that performs case-based active and interactive learning to discover goal selection knowledge. We also report tests of its performance in a complex environment. We found that, under some conditions, T-ARTUE can quickly learn goal selection knowledge.

The combination of learning by demonstration and planning has proved an effective solution for real-time strategy games. Nevertheless, learning hierarchical plans from expert traces also has its limitations regarding the number of training traces required, and the absence of mechanisms for rapidly reacting to high priority goals. We propose to bring the game designer back into the loop, by allowing him to explicitly inject decision making knowledge, in the form of behavior trees, to complement the knowledge obtained from the traces. By providing a natural mechanism for designers to inject knowledge into the plan library, we intend to integrate the best of both worlds: learning from traces and hard-coded rules.

A parallel plan is a sequence of sets of actions such that any ordering of actions in the sets gives a traditional sequential plan. Parallel planning was popularized by the Graphplan algorithm and it is one of the key components of successful SAT-based planers. SAT-based planners have recently begun to exploit multi-valued state variables – an area which seems traditionally more suited for constraint-based planners – and they improved their performance further. In this paper we propose a novel view of constraint-based planning that uses parallel plans and multi-valued state variables. Rather than starting with the planning graph structure like other parallel planners, this novel approach is based on the idea of timelines and their synchronization.

Current evaluation functions for heuristic planning are expensive to compute. In numerous planning problems these functions provide good guidance to the solution, so they are worth the expense. However, when evaluation functions are misguiding or when planning problems are large enough, lots of node evaluations must be computed, which severely limits the scalability of heuristic planners. In this paper, we present a novel solution for reducing node evaluations in heuristic planning based on machine learning. Particularly, we define the task of learning search control for heuristic planning as a relational classification task, and we use an off-the-shelf relational classification tool to address this learning task. Our relational classification task captures the preferred action to select in the different planning contexts of a specific planning domain. These planning contexts are defined by the set of helpful actions of the current state, the goals remaining to be achieved, and the static predicates of the planning task. This paper shows two methods for guiding the search of a heuristic planner with the learned classifiers. The first one consists of using the resulting classifier as an action policy. The second one consists of applying the classifier to generate lookahead states within a Best First Search algorithm. Experiments over a variety of domains reveal that our heuristic planner using the learned classifiers solves larger problems than state-of-the-art planners.

Most current planners assume complete domain models and focus on generating correct plans. Unfortunately, domain modeling is a laborious and error-prone task. While domain experts cannot guarantee completeness, often they are able to circumscribe the incompleteness of the model by providing annotations as to which parts of the domain model may be incomplete. In such cases, the goal should be to generate plans that are robust with respect to any known incompleteness of the domain. In this paper, we first introduce annotations expressing the knowledge of the domain incompleteness, and formalize the notion of plan robustness with respect to an incomplete domain model. We then propose an approach to compiling the problem of finding robust plans to the conformant probabilistic planning problem.

Sequential decision tasks present many opportunities for the study of transfer learning. A principal one among them is the existence of multiple domains that share the same underlying causal structure for actions. We describe an approach that exploits this shared causal structure to discover a hierarchical task structure in a source domain, which in turn speeds up learning of task execution knowledge in a new target domain. Our approach is theoretically justified and compares favorably to manually designed task hierarchies in learning efficiency in the target domain. We demonstrate that causally motivated task hierarchies transfer more robustly than other kinds of detailed knowledge that depend on the idiosyncrasies of the source domain and are hence less transferable.

David McAllester and David Rosenblitt's paper, "Systematic Nonlinear Planning" (published This commentary by Daniel S. Weld describes David Rosenblitt's paper, "Systematic Nonlinear Planning" (McAllester and Rosenblitt 1991), presented 19 years ago at the Ninth National Conference on Artificial Intelligence (AAAI-91), had two major impacts on the field: (1) an elegant algorithm and (2) endorsement of the lifting technique. The paper's biggest impact stems from its extremely clear and simple presentation of a sound and complete algorithm (known as SNLP or POP) for classical planning. While it is easy to define such an algorithm as search through the space of world states, SNLP is a "partialorder" planner, meaning it searches the space of partially specified plans, where only partial constraints on action arguments and ordering decisions are maintained. Here, McAllester and Rosenblitt benefited from David Chapman's elegant TWEAK planner, which greatly clarified previous partial-order algorithms (Chapman 1985). SNLP's key feature is the use of a data structure, called a causal link, to record the planner's commitment to establish a precondition of one action with the postcondition of another.

We have explored the use of analogy as a general approach to near and far transfer learning in domains ranging from physics problem solving to strategy games (Klenk and Forbus 2007; Hinrichs and Forbus 2007). Using the same basic analogical mechanism, we have found that the main differences between near and far transfer involve the amount of generalization that must be performed prior to transfer and the way that the matching process treats nonidentical predicates. We present here two extensions of our analogical matcher, minimal ascension and metamapping, that enable far transfer between representations with different relational vocabulary. Evidence for the effectiveness of these techniques is provided by a large-scale external evaluation, involving a substantial number of novel distant analogs.

Virtual and robotic agents must be able to understand "communicative acts" (utterances, gestures, controlled facial expressions etc.) if they are to interact and collaborate with humans. For researchers in AI, HCI, HRI and related fields, automatic comprehension of communicative acts has turned out to be a very tough nut to crack. Drawing on recent research from cognitive science and evolutionary psychology, the paper argues that an insufficient conceptualization of "context" is at the heart of this problem, and that we should focus on very simple, non-linguistic communicative acts (pointing gestures etc.) in order to investigate how agents can comprehend communicative acts in realistic contexts. I propose a tripartite model of context which is informed by experimental research on how humans recognize objects (via "affordances"), causal relations among objects, and the collaborative activities of fellow-humans. The model is not a formal one, but detailed enough to help in the development of comprehension algorithms in future research.