In this paper we examine the behavior of a human-computer system for crisis response. As one instance of crisis management, we describe the task of responding to spills and fires involving hazardous materials. We then describe Inca, an intelligent assistant for planning and scheduling in this domain, and its relation to human users. We focus on Inca’s strategy of retrieving a case from a case library, seeding the initial schedule, and then helping the user adapt this seed. We also present three hypotheses about the behavior of this mixed-initiative system and some experiments designed to test them. The results suggest that our approach leads to faster response development than user-generated or automatically-generated schedules but without sacrificing solution quality.

Evolutionary algorithms search for problem solutions by selecting individuals for survival and reproduction with a bias towards higher fitness. Such biases may lead to premature convergence on sub-optimal solutions. A bias toward greater diversity can help delay convergence and broaden the area searched for optimal candidate solutions. We introduce two ways to measure a population's diversity and evaluate how they interact with traditional fitness during selection. We then introduce a mechanism that includes a bias toward greater diversity in addition to traditional accuracy. Using the King-Rook-King chess endgame problem, we demonstrate that including diversity as a selection factor leads to better overall solutions.

Fictional stories abound where artificial intelligence gives rise to man-made servants that make our lives a holiday (or a nightmare). Naturally, the optimist has a strong motive to ask: How might we design and implement an artificial servant? We regularly operate under the notion that one agent helps another by doing something for the other. However, the story behind this is much more complicated. In this position paper, we explore two questions: What is the nature of service that is helpful or “good”? and What will it take to design and implement agents that provide “good” service? We start by describing our current understanding of the nature of service, consisting of a number of dimensions along which service may be measured and compared, and three modes in which service may be delivered. With these dimensions and modes setting the stage, we consider the technical and social requirements for implementing artificial servants. We find that while many prerequisites are either already available or could reasonably be developed from the current state of the art, both technical and social challenges remain that are expected to be very difficult to overcome. Our experimental work attempts to generate and evaluate the fundamental elements of helpful assistance, and we provide a brief overview of our work. One surprising empirical result reminds us of a common understanding about help: providing assistance is a two-way street. Developing that two-way street may prove to be one of the most significant obstacles to realizing an artificial personal servant.

For many classification problems, genetic algorithms prove to be effective without extensive domain engineering. However, the chess King-Rook-King endgame problem appears to be an exception. We explore whether modifications to a baseline parallel genetic algorithm can improve the accuracy on this particular problem. After describing the problem domain and our implementation of a parallel genetic algorithm, we present an empirical evaluation of several approaches intended to improve overall performance. Our results confirm the challenging nature of this domain. We describe several directions that may yet deliver significant improvements.

Movies such asThe Matrixhave stimulated popular interest in “brain in a vat” scenarios. Amidst the traditional questions of mind, we tend to overlook an integral enabling component – the world simulation – which merits consideration in its own right. When facing the simulations in these imagined scenarios, we struggle with conceptual muddles regarding what is “real” and not. In this paper, I argue that simulated worlds are every bit as real as the one we inhabit. This turns out to be important when considering the possibility, as suggested by Nick Bostrom (2003), that the world we experience as “real” is actually a simulation. Can such a hypothetical prospect be reconciled with an orthodox Christian perspective? While the metaphysical status of simulations that I present here points towards such a reconciliation, significant obstacles remain to be addressed. I consider s om of these remaining challenges and explore the associated stakes. Although this consideration works through a hypothetical scenario, the exercise p rovides several insights that may prove valuable apart from the required a s-

Exploring Moral Reasoning in a Cognitive Architecture Wayne Iba Pat Langley Computer Science Department Westmont College, 955 La Paz Road Santa Barbara, CA 93108 Institute for the Study of Learning and Expertise 2164 Staunton Court, Palo Alto, CA 94306 Abstract Background Moral reasoning plays a significant but poorly under- stood role in human action and interaction. Although studied by philosophers for millennia, considerable con- fusion surrounds the topic. Computational cognitive ar- chitectures hold promise for shedding insight on how agents act and reason morally. We present a view of moral cognition and examine one implementation of that view in Icarus, a theory of the human cognitive archi- tecture. This approach to moral behavior and reasoning leads us to suggest that morality is a special case of ev- eryday cognition. We discuss the implications of this view and outline our continuing research on these and related questions. Keywords: Machine ethics; Moral reasoning; Philoso- phy; Cognitive architectures. In this paper, we describe our attempt to model moral cognition within an integrated cognitive architecture. As such, we draw upon work in three distinct disciplines – philosophy, psychology and computer science – each of which has its own body of literature. In this section we point to a very small sample of work that has influenced our thinking. Although a distinction may be made between ethics and morals, for our present purposes we will use the terms interchangeably. The point of an ethical theory is to provide guidelines for how to live “the good life”. As such, an ethical theory should identify the nature of the good as well as provide a system for acting so as to achieve the good. Here we are not advancing a particular theory, but rather exploring how a specific cognitive architecture could support moral reasoning of different types. However, it will help readers to have at least a basic familiarity of the main ethical theories for discussions in later sections. The three dominant ethical theories and their primary initiators include the virtue ethics of Aristotle, deonto- logical systems descending from Kant, and consequen- tialist theories following from Hume, Bentham, and Mill. In virtue ethics, one lives the good life by behaving in a balanced manner, thereby exhibiting the virtues – nei- ther too much nor too little of a given quality. Deon- tological systems place value on fulfilling one’s duties or obligations; commonly, this is viewed as following cer- tain rules that prescribe how one ought to behave. Con- sequentialist theories assign different values to different states and prescribe courses of action that result in states with the best values; utilitarianism is one particular form of consequentialism where actions are selected so as to maximize the total good across the entire population. One approach to understanding moral reasoning is to observe and explain what people do when they are faced with morally charged decisions (Baron & Ritov, 2009; Waldmann & Dieterich, 2007). In contrast to the philosophical theories that prescriptively indicate what “ought” to be done, this approach attempts to charac- terize what actually takes place and why. Such studies provide hints and constraints on determining the nature of moral decision making in humans (Spranca, Minsk, & Baron, 1991). For example, many of these studies sug- gest that people operate using a hybrid of deontic and Introduction Moral reasoning has been a focus of study for millennia. The human condition stems in large part from the col- lective judgments and decisions that can be said to have moral valence. This suggests that it would be desir- able to study this topic carefully. Nevertheless, our un- derstanding of morally charged cognition is still incom- plete and uncertain. Since the introduction of comput- ers, many people have been fascinated with the prospect of writing programs that exhibit human intelligence. As with other aspects of intelligence, computational mod- els can provide many insights into the nature of moral cognition. Although generally treated as a topic distinct from generic reasoning, we suggest here that moral cogni- tion is better viewed as a special case of practical cog- nition. Depending on the ethical approach one takes, moral dilemmas might be nothing other than the conse- quence of bounded rationality. Perhaps our subjective experience when deliberating over a moral dilemma has more to do with an emotion than with the properties of the problem that we are trying solve. These two factors – bounded rationality and emotional states – may be the actual source of moral dilemmas. We start by distinguishing between three types of moral cognition: moral behavior, moral interpretation, and moral decision making. Next we review Icarus, the cognitive architecture we have used to implement agents that exhibit forms of moral cognition. We discuss our experiences with these agents and their broader implica- tions for accounts of this class of phenomena. In closing, we suggest directions for future research that should fur- ther clarify our understanding of moral cognition.

Norms or conventions can be used as external correlating signals to promote coordination between rational agents and hence have merited in-depth study of the evolution and economics of norms both in the social sciences and in multiagent systems. While agent simulations can be used to gain a cursory idea of when and what norms can evolve, the estimations obtained by running simulations can be costly to obtain, provide no guarantees about the behavior of a system, and may overlook some rare occurrences. We use a theoretical approach to analyze a system of agents playing a convergence game and develop models that predict (a) how the system's behavior will change over time, (b) how much time it will take for it to converge to a stable state, and (c) how often the system will converge to a particular norm.

In many multi-agent systems, the emergence of norms is the primary factor that determines over-all behavior and utility. Agent simulations can be used to predict and study the development of these norms. However, a large number of simulations is usually required to provide an accurate depiction of the agents' behavior, and some rare contingencies may still be overlooked completely. The cost and risk involved with agent simulations can be reduced by analyzing a system theoretically and producing models of its behavior. We use such a theoretical approach to examine the dynamics of a population of agents playing a coordination game to determine all the norms to which the society can converge, and develop a system of linear recurrence relations that predict how frequently each of these norms will be reached, as well as the average convergence time. This analysis produces certain guarantees about system behavior that canot be provided by a purely empirical approach, and can be used to make predictions about the emergence of norms that numerically match those obtained through large-scale simulations.

Classifying the endgame positions in Chess can be challenging for humans and is known to be a difficult task in machine learning. An evolutionary algorithm would seem to be the ideal choice. We describe our implementation of a parallel island model and evaluate it in the context of the Chess Endgame data set from the UCI machine learning repository. We are mainly interested in impact of parallelization upon runtime and accuracy. Thus, we compare the system’s performance under a number of varied conditions, including population size, number of islands, number of neighbors, migration rate, and migrant selection strategy. These results show the system to be useful from an efficiency standpoint, and point to opportunities to better understand the behavior and properties of the islandmodel evolutionary algorithm.

Problems and methods encountered in artificial intelligence hold great promise for inspiring interest in computer science in general and AI in particular. The challenge faced in using such problems in introductory computer science courses is three-fold: to prune the complexity that is beyond beginning students, but to preserve the engaging nature of the problems, and still exercise the topics or skills from the beginning courses. In this paper, I present a number of extended exercises that attempt to balance those three challenges. These projects have been used multiple times in CS1 and CS2 courses; the projects include: animated swarming of multiple critter types, cellular automata, neural networks, searching game trees, and a simple genetic algorithm. Lessons learned from employing these projects, drawn from personal reflection and anecdotal reports from students, suggests using AI problems benefits both the students and the instructor.

Crisis domains present the challenge of developing good responses in a timely manner. In this paper, we present an interactive, case-based approach to crisis response that provides users with the ability to rapidly develop good responses while leaving ultimate decision-making control to the users. We introduce Inca, the INteractive Crisis Assistant we have implemented for planning and scheduling in crisis domains. We also present Haz-Mat, the artiicial domain involving hazardous material incidents that we developed for the purpose of evaluating diierent responses and various assistant mechanisms. We then discuss two preliminary studies that we conducted to evaluate scheduling assistance in Inca. Results from the rst set of experiments indicate that Inca's case-based scheduling assistance provides users with initial candidate solutions that enable users to develop high quality responses more quickly. The second set of experiments demonstrates the potential of machine learning methods to further facilitate interactive scheduling by accurately predicting preferred user adaptations. Based on these encouraging results, we close with directions for future work and a brief discussion of related research.

We present a computational model of movement skill learning. The types of skills addressed are a class of trajectory following movements involving multiple accelerations , decelerations and changes in direction and lasting more than a few seconds. These skills are acquired through observation and improved through practice. We also review the speed-accuracy tradeoo|one of the most robust phenomena in human motor behavior. We present two speed-accuracy tradeoo experiments where the model's performance ts human behavior quite well.

We regularly operate under the notion that one agent assists another when the first does something for the second. However, the story behind this is much more complicated. In this position paper, we explore two questions: How can we evaluate the quality or goodness of a particular instance of assistance? and How can we design agents to initiate and provide “good” assistance? In asking these questions, we are assuming two things. First, performing a task for another agent is not always helpful to that agent even if the task really needs to be done. But second, a helper that does not have the skill or resources to accomplish a specifically requested task may still provide real help. Our experimental work attempts to address the fundamental elements of helpful assistance. One surprising initial result reminds us of an common understanding about help: providing assistance is a two-way street. Measuring the Helpfulness of Assistance. Providing assistance can be one of those activities that is performed for the benefit of another agent without sufficient reflection. We want to better understand the nature of assistance that is truly helpful. At one point or another, everyone has experienced service that was exceptional or memorable – either because it was unusually good or unusually bad. Either at a restaurant, an auto mechanic, or a health clinic, we have been “assisted” by someone who was inattentive, incompetent, unscrupulous, or even all three at the same time. Likewise, most of us have encountered situations where the service provider anticipated our needs, took care of those needs efficiently, or treated us with respect and courtesy. This common experience forms the basis of my primary claim, which we consider to be self-evident: Assumption 1: All assistance is not created equal. Or in other words, not all assistance is actually helpful. If we are willing to grant this assumption, then at least two questions should immediately spring to mind: How can we evaluate the quality of assistance? and How can we construct ∗We are grateful to Westmont Provost, Shirley Mullen, and the Westmont Professional Development Committee. Copyright c © 2007, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. agents that truly are helpful? Both questions have implicitly been at the core of essentially all work in the area of intelligent assistants. However, the implicit answers to the first question are sometimes overly narrow. Question 1: How can we evaluate the quality or goodness of assistance? A traditional answer to this question focuses on the assistant alone and consists of specifying a task that needs to be done and then measuring how often or how completely an assistant accomplishes the given task. Essentially, the approach measures the competence of a service provider’s assistance in a given encounter. Unfortunately, this addresses only a single facet of assistance. We identify at least five other dimensions that influenced the positive or negative assessment of an interaction. Other dimensions along which we can view an assistant include: attention, anticipation, persistence deference and integrity. Briefly, each of these other dimensions allow us to describe characteristics of assistance that contribute to our assessment of a particular engagement as being helpful or not. For instance, we expect an assistant to pay attention to our current situation and track changes in our needs over time. We would also hope that our assistant can anticipate our needs and satisfy those needs or perform tasks that contribute to our goals in a timely fashion and perhaps without being asked.1 Once given a task, we hope that an assistant will be persistent in working on and completing the task. Also, an assistant should operate with some deference toward those it is intending to assist; this might amount to giving the goals of the recipient a higher priority than the goals of the assistant itself. Finally, an assistant should operate with integrity; that is, information and services must not be misrepresented and payments charged for the assistance should be fair. Each of these, together with the competence of an assistant, combine to determine the overall helpfulness of the interaction. Reflection on the dimensions suggests at least two corollary assumptions to our central claim. These assumptions may provide further constraints on the design of assistants. In this context, we do not wish to address the thorny issue of autonomy or initiative; this certainly requires additional consideration. Minimally, we can say that accurate anticipation is better than faulty anticipation whether or not the assistant is able to act independently on that anticipation. Assumption 1.1: Competent helpers are not enough. Good assistance requires more than simply performing a task for another agent. This is true even when the task in question really needs to be done. For example, if two subtasks must be completed in a specific sequence, then an assistant that arbitrarily does one task may not be helping. Or two sub-tasks may have significantly different difficulties and help is needed with the most difficult one; in this case, an assistant that performs the easier task may not be as helpful as the one that works on the more difficult task. Again, performing a task may bring a user enjoyment and having the task completed by an assistant could deprive the user of that satisfaction. Thus, we have several examples where an assistant may be competent, anticipatory, and persistent, yet may fail to be helpful. Note then that these features – competence, anticipation, and persistence – are not sufficient for helpful assistance. But if the examples above are not helpful, what is missing? In each case, an assistant that was attentive to the goals of the one receiving the help would have avoided the problems identified. That is not to say that attention is then the key to good service; other examples involving attentive but incompetent assistants would demonstrate the nonsufficiency of attention. To sharpen this point, let us consider situations where we might be surprised to find helpful assistance. Assumption 1.2: Incompetent helpers can be helpful. The long tradition of apprenticeship provides adequate evidence that assistants with less than masterful skills can provide helpful service while they are developing their expertise. A child helping its parent on a project around the house can also provide real help by fetching tools or parts. Together with the previous assumption, this suggests that competence is neither a necessary nor sufficient feature of good assistance. We claim that there are no set of necessary and sufficient features that define “good” assistance. If we want to carefully and fully evaluate the quality of service delivered by our intelligent assistants, then we need to consider all of these dimensions. However, in addition to providing a strategy for evaluating assistants, these dimensions also suggest behaviors or inclinations that assistants should display, and thus contribute to an answer to our second question. Question 2: How do we design agents that can deliver “good” assistance? Of course, we ultimately want to design and implement assistants that can be truly helpful. The dimensions identified above may be generally useful for evaluating the assistance provided by a given agent compared to that of another. They may also suggest designs for how to implement assistants. Most simply, an assistant design should respond to each of the six dimensions. However, because they are neither necessary nor sufficient, assistants without one or more of these capabilities may still prove helpful. Nevertheless, holding all other things equal, an improvement along one dimension leads to an overall improvement in helpfulness. Empirical Work. We have conducted our work within the context of the MÆDEN simulated environment (Iba & Burwell 2005a; 2005b). This multi-agent environment supports a variety of tasks, agent architectures and communication protocols. Using the GARCIA agent framework(Iba & Holm 2006), we have implemented agents that ask for and deliver assistance in the midst of problem solving tasks. The agent framework addresses all of the dimensions discussed above; specifically, we can vary competence, attention, anticipation, persistence, deference and integrity. Our previous experiments have addressed competence, persistence, and integrity. We evaluated problem solving success rates and efficiency while varying the skills that an assistant possessed. As expected, our dependent measures (number of problems solved and average cost to solve them) improved with increasing skills. We also varied parameters that control persistence and integrity with the obvious anticipated results. So far, we have mostly focused on an incompetent agent seeking help from an assistant that is more skilled to varying degrees. A somewhat surprising result that emerged from our previous work highlights the role of the recipient in the exchange of assistance. Based on a level of patience, an agent may request help but later give up on the assistant and continue trying to solve the problem on its own (possibly asking for help again later). We observed that the agent’s commitment to the assistant positively influenced the overall success rate and efficiency. Although it is commonly understood that unteachable students cannot be helped, we were not expecting to encounter this effect in our tests. Thus, we intend to extend our dimensions to include factors pertaining to the agent receiving assistance. At the very least, we can say the quality of the assistance provided will depend on the recipient’s willingness to accept and integrate the assistant’s work. However, we suspect that there are other dimensions that pertain to the service recipient that need to be identified. Along these lines, our current experiments vary the competence of both the

Assistance, or service, is becoming increasingly important in the information economy, but few people have studied it. MAEDEN provides a testbed for studying service using a variety of agent architectures. We are developing GARCIA, an architecture that supports interactions between agents and where one agent can provides assistance to another that is in need of help. We present several recent empirical studies that go beyond previous work with the intent of identifying fundamental features

Performing a task for a user is one part of assistance. As a prior step, a helper must identify which of a possible number of tasks should be assisted. We intend to study the requirements that enable service to be rendered and understand the factors that impact the delivery of assistance. Toward this end, we extended a simulated environment to use as our testbed for developing and evaluating approaches to service. Our new testbed, M˘DEN, supports multiple agents, communication between agents, and records detailed logs of agents’ behavior. We used the testbed in preliminary studies of servant and recipient behavior and interaction. Results hint at the usefulness of the testbed.

If we are to provide systems that benet human users, we need a better understanding of the nature of service. In this paper, we describe the necessary characteristics of a simulator for studying service, and then present M˘DEN, our attempt to satisfy those requirements. We then proceed to describe an initial set of agents that we use to explore the nature of service. Our experiments conrm the usefulness of M˘DEN as a testbed, but also point toward insights into service in general. We end with a discussion of future work intended to explore those insights. Overview In this paper, we address the problem of service. Because the nature of service is difcult to understand, we face particular challenges when trying to build systems that provide valuable service to users. In this section, we describe our view of service and review previous relevant work. Then in the following section, we describe our work extending a world simulator to support the study of service. Next, we present our design and implementation of agents that give and receive service. We follow this with the results of our rst experiments and conclude with a discussion of future work and lessons learned.

Computer systems are becoming extremely powerful as well as more pervasive. In order to maximize their benefits, we want to provide adaptive interface layers between these systems and their users. Specifically, we expect the interface layer to adapt to unique characteristics of a particular user. Ultimately, we hope to uncover general principles of adaptive user interface design, which include techniques for modeling user habits, gathering information and feedback to drive those models, and a methodology for quantifying the value embodied in a particular adaptive user interface. Toward these ends, we have explored several problem domains and several modeling techniques. We have tested some of the combinations on users and are in the process of designing methods to test the others. Our strategy has been to first focus on the hardest problems with the assumption that we could learn the most from these. Thus, we have focused primarily on the crisis response domain, a domain where the user (responder) must cope with threat, uncertainty, and urgency. In this short abstract, we sketch our research with respect to crisis response, and then close with a summary of our efforts to generalize our methods and results.

Since the field’s inception, most research in machine learning has focused on the problem of supervised induction from labeled training cases. If anything, this trend has been strengthened by the creation of data repositories that, typically, include class information. But this emphasis is misguided if we want to understand the nature of learning in intelligent agents like humans. Clearly, children acquire many concepts about the world before they learn names for them, and scientists regularly discover patterns without any clear supervision from an outside source. Even the availability of class labels in public data sets can be misleading; many such domains are medical in nature, and medical researchers first had to discover a disease before they could diagnose it for particular patients.