However, if it is incorrect, there must be some relationship between the two that allows them to live harmoniously within a single theory. The nature of this relationship is the focus of this article. In passing we note that the theory of weak problem solvers has been well-developed for over a decade; see Kilsson (1971) for example. Some aspects of MYCIN don't fit the problem reduction For example, a THE AI MAGAZINE Summer 1983 25 production whose action part is a conjunction of atomic formulae corresponds to a separate operator for each atomic formula in the conjunction. MYCIN's search strategy effectively applies such operators in a group.

What Do We Know about Knowledge? However, the simple equation "knowledge is power" leaves three major questions unanswered. First, what do we mean by "knowledge"; second, what do we mean by "power"; and third, what do we mean by "is"? In this article, I will examine the first of these questions. In particular I will focus on some of the milestones in understanding the nature of knowledge and some of what we have learned from 50 years of AI research. The discipline and detail required to write programs that use knowledge have given us some valuable lessons for implementing the knowledge principle, one of which is to make our programs as flexible as we can. Many of them are well known, but they can serve as reminders of the difficulty of implementing the "knowledge is power" principle. I wish to clarify the knowledge principle and try to increase our understanding of what programmers and program designers need to do to make the knowledge principle work in practice. The "knowledge is power" principle is most closely associated with Francis Bacon, from his 1597 tract on heresies: "Nam et ipsa scientia potestas est." ("In and of itself, knowledge is power.") Bacon was among the first of the modern philosophers to separate the concept of scientific knowledge from knowledge gained through the two dominant methods for attaining truth in his time: magic and religious revelation. The essential difference for him, as for us, is that knowledge gained through experiment is replicable by others. Although all the empirical sciences rely on the replication of observations and experiments, AI has been slow to embrace this principle. Programs demonstrating research ideas in AI are often too large and not well enough documented to allow replication or sharing. Applications programs, however, are designed to be used by others outside the research lab and thus are more amenable to multiple runs in diverse conditions. Thus they have the potential to provide experimental data demonstrating strengths, weaknesses, and benefits. The knowledge principle predates Bacon. For example, it was pretty clearly articulated in Biblical times: "A man of knowledge increaseth strength" (Proverbs 24: 5). Socrates, Plato, Aristotle, and other early Greek philosophers based their lives on acquiring and transferring knowledge. In the course of teaching, they sought to understand the nature of knowledge and how we can establish knowledge of the natural world. Socrates is famous for pointing out the value of knowledge and seeking truth, as in "… that which we desire to have, and to impart to oth-

Research on mixed-initiative interaction and assistance is still in its infancy but is poised to blossom into a wellspring of innovation that promise to change the way we work with computing systems--and the way that computing systems work with us. I share reflections about the opportunities ahead for developing computational systems with the ability to engage people in a deeply collaborative manner, founded on their ability to support fluid mixed-initiative problem solving. Such collaborative intelligence sits at the veritable heart of human civilization. In the course of daily life, we assume and rely on a rich interleaving of efforts to achieve goals while immersed in shared context. We continue to engage one another in efficient, tightly woven collaborations, reasoning with remarkable efficiency about the beliefs, preferences, intentions, and skills of potential collaborators. The inferences underlying successful collaborations typically stream in such an effortless and subconscious manner that we often fail to recognize the elegance and sophistication of these capabilities. The magic of human collaborative competency comes to the foreground with attempts to extend these skills to computational systems. Developing a better understanding of the core aspects of intelligence that enable people to collaborate with fluidity promises to enable new kinds of human-computer collaboration. The nascent area of research on mixed-initiative interaction centers on developing methods that enable computing systems to support an efficient, natural interleaving of contributions by people and computers, aimed at converging on solutions to problems. In mixed-initiative interaction, people and computers take initiatives to contribute to solving a problem, achieving a goal, or coming to a joint understanding. Conversational dialogue is an oft-cited example of mixed-initiative interaction, referring to the ability of each participant in a dialogue to take initiative to guide or add to a discussion. Endowing an automated dialogue system with the ability both to take initiative ("What city do you wish a flight to?") and to allow people to take conversational initiative ("Wait, I'd like to add a side trip.") However, mixed-initiative interaction extends beyond spoken conversations to include a broad spectrum of collaborative problem solving marked by an interleaving of contributions by different participants. Mastering mixed-initiative interaction poses a constellation of fascinating challenges and opportunities for AI researchers. Figure 1 highlights the core challenge of seeking mutual understanding or grounding of joint activity. Joint activity describes the behavior displayed by people working together to solve a mutual goal.

The fourth Modeling and Reasoning in Context (MRC) workshop was held on August 20-21, 2007, in conjunction with the Sixth International and Interdisciplinary Conference on Modeling and Using Context, at Roskilde University, Denmark. This year's workshop included a special track on the role of contextualization in human tasks (CHUT). The overall goal of the workshop was to further the understanding, development, and application of AI methods for context-sensitive information technology. The Modeling and Reasoning in Context (MRC) workshop series, begun in 2004, brings together researchers and practitioners to exchange ideas and results on modeling and reasoning issues for context-sensitive systems. MRC 2007 broadened the focus to also highlight studies of contextualization in human tasks (CHUT), to explore the practical relationships between tasks, actors, and workplace context that may shape system design. The workshop was split into formal paper presentations and discussion sessions. The first two discussions combined themed panels with audience participation, while the closing free-form discussion offered the opportunity for participants to examine issues of their choice and provide closing perspective on the workshop as a whole. Following an MRC tradition, the workshop also included an informal dinner, enabling participants to continue their discussions in a traditional Copenhagen restaurant. The MRC paper presentations covered topics such as ontology-based context models, the benefits of multilayered models (combining general metalevel and domain models with applicationspecific instances), the use of situation lattices to achieve situation awareness, user modeling in mobile ambient intelligent systems, and middleware for managing context. These were illustrated for a range of tasks, such as contextualized software reuse and an email filtering approach using multiple heterogeneous sources of contextual data to infer when and where to deliver messages. The contextualization of human tasks was demonstrated from multiple perspectives as well, ranging from analysis of interpersonal work practices, to discover contextual parameters, to an application to improve drivers' situation awareness. These diverse presentations gave a good overview of the various uses of context, their benefits, and their challenges for modeling and reasoning, providing a starting point for the discussions. There was enthusiastic participation in the workshop's discussions, and many participants considered the exchanges there to be the most rewarding part of the workshop.

Automated planning is a branch of AI that addresses the problem of generating a set of actions to achieve a specified goal state, given an initial state of the world. It is an active area of research that is central to the development of intelligent agents and au - tonomous robots. In many real-world applications, a multitude of valid plans exist, and a user distinguishes plans of high quality by how well they adhere to the user's preferences. To generate such high-quality plans automatically, a planning system must provide a means of specifying the user's preferences with respect to the planning task, as well as a means of generating plans that ideally optimize these preferences. In the last few years, there has been significant research in the area of planning with preferences.

In the winter 2004 issue of AI Magazine, we reported Vulcan Inc.'s first step toward creating a question-answering system called Digital Aristotle. The goal of that first step was to assess the state of the art in applied knowledge representation and reasoning (KRR) by asking AI experts to represent 70 pages from the advanced placement (AP) chemistry syllabus and to deliver knowledge-based systems capable of answering questions from that syllabus. This article reports the next step toward realizing a Digital Aristotle: we present the design and evaluation results for a system called AURA, which enables domain experts in physics, chemistry, and biology to author a knowledge base and that then allows a different set of users to ask novel questions against that knowledge base. These results represent a substantial advance over what we reported in 2004, both in the breadth of covered subjects and in the provision of sophisticated technologies in knowledge representation and reasoning, natural language processing, and question answering to domain experts and novice users. As this capability develops, the project focuses on two primary applications: a tutor capable of instructing and assessing students and a research assistant with the broad, interdisciplinary skills needed to help scientists in their work.

More specifically, we demonstrate how to transfer the domain-dependent heuristics acquired by one planner into a second planner. Our motivation is to improve the efficiency and the efficacy of the second planner by allowing it to use the transferred heuristics to capture domain regularities that it would not otherwise recognize. Our experimental results show that the transferred knowledge does improve the second planner's performance on novel tasks over a set of seven benchmark planning domains. Recently, the artificial intelligence community has attempted to model this transfer in an effort to improve learning on new tasks by using knowledge from related tasks. For example, classification and inference algorithms have been extended to support transfer of conceptual knowledge (for a survey see Torrey and Shavlik [2009]).

The establishment of the International General Game Playing Competition in 2005, however, resulted in a renewed interest in more general problem-solving approaches to game playing. In general game playing (GGP) the goal is to create gameplaying systems that autonomously learn how to play a wide variety of games skillfully, given only the descriptions of the game rules. In this paper we review the history of the competition, discuss progress made so far, and list outstanding research challenges. Unlike specialized game players, such as Deep Blue (Campbell, Hoane, and Hsu 2002), general game players cannot rely on algorithms designed in advance for specific games; they must discover such algorithms themselves. General game playing expertise depends on intelligence on the part of the game player rather than intelligence of the programmer of the game player.

Useful feedback makes use of models of domain-specific knowledge, especially models that are commonly held by potential students. To empirically determine what these models are, student data can be clustered to reveal common misconceptions or common problem-solving strategies. This article describes how analogical retrieval and generalization can be used to cluster automatically analyzed handdrawn sketches incorporating both spatial and conceptual information. We use this approach to cluster a corpus of hand-drawn student sketches to discover common answers. Common answer clusters can be used for the design of targeted feedback and for assessment.

The aim of the Angry Birds AI competition (AIBIRDS) is to build intelligent agents that can play new Angry Birds levels better than the best human players. This is surprisingly difficult for AI as it requires similar capabilities to what humans need for successfully interacting with the physical world, one of the grand challenges of AI. As such the competition offers a simplified and controlled environment for developing and testing the necessary AI technologies, a seamless integration of computer vision, machine learning, knowledge representation and reasoning, reasoning under uncertainty, planning, and heuristic search, among others. Over the past three years there have been significant improvements, but we are still a long way from reaching the ultimate aim, and thus, there are great opportunities for participants in this competition. The competition was initiated in 2012 by the authors of this report and is held in collocation with some of the major AI conferences such as the International Joint Conference on Artificial Intelligence in 2013 and again in 2015, and the European Conference on Artificial Intelligence conference in 2014.