We introduce a new algorithm for off-policy temporal-difference learning withfunction approximation that has lower variance and requires less knowledge of the behavior policy than prior methods. We develop the notion ofa recognizer, a filter on actions that distorts the behavior policy to produce a related target policy with low-variance importance-sampling corrections. We also consider target policies that are deviations from the state distribution of the behavior policy, such as potential temporally abstract options, which further reduces variance. This paper introduces recognizers and their potential advantages, then develops a full algorithm for linear function approximation and proves that its updates are in the same direction as on-policy TD updates, which implies asymptotic convergence. Eventhough our algorithm is based on importance sampling, we prove that it requires absolutely no knowledge of the behavior policy for the case of state-aggregation function approximators.

This paper explores two aspects of social network modeling. First, we generalize a successful static model of relationships into a dynamic model that accounts for friendships drifting over time. Second, we show how to make it tractable to learn such models from data, even as the number of entities n gets large.

In real-life temporal scenarios, uncertainty and preferences are often essential and coexisting aspects. We present a formalism where quantitative temporal constraints with both preferences and uncertainty can be defined. We show how three classical notions of controllability (that is, strong, weak, and dynamic), which have been developed for uncertain temporal problems, can be generalized to handle preferences as well. After defining this general framework, we focus on problems where preferences follow the fuzzy approach, and with properties that assure tractability. For such problems, we propose algorithms to check the presence of the controllability properties. In particular, we show that in such a setting dealing simultaneously with preferences and uncertainty does not increase the complexity of controllability testing. We also develop a dynamic execution algorithm, of polynomial complexity, that produces temporal plans under uncertainty that are optimal with respect to fuzzy preferences.

The best performing algorithms for a particular oversubscribed scheduling application, Air Force Satellite Control Network (AFSCN) scheduling, appear to have little in common. Yet, through careful experimentation and modeling of performance in real problem instances, we can relate characteristics of the best algorithms to characteristics of the application. In particular, we find that plateaus dominate the search spaces (thus favoring algorithms that make larger changes to solutions) and that some randomization in exploration is critical to good performance (due to the lack of gradient information on the plateaus). Based on our explanations of algorithm performance, we develop a new algorithm that combines characteristics of the best performers; the new algorithm's performance is better than the previous best. We show how hypothesis driven experimentation and search modeling can both explain algorithm performance and motivate the design of a new algorithm.

It was recently proved that a sound and complete qualitative simulator does not exist, that is, as long as the input-output vocabulary of the state-of-the-art QSIM algorithm is used, there will always be input models which cause any simulator with a coverage guarantee to make spurious predictions in its output. In this paper, we examine whether a meaningfully expressive restriction of this vocabulary is possible so that one can build a simulator with both the soundness and completeness properties. We prove several negative results: All sound qualitative simulators, employing subsets of the QSIM representation which retain the operating region transition feature, and support at least the addition and constancy constraints, are shown to be inherently incomplete. Even when the simulations are restricted to run in a single operating region, a constraint vocabulary containing just the addition, constancy, derivative, and multiplication relations makes the construction of sound and complete qualitative simulators impossible.

Allocating scarce resources among agents to maximize global utility is, in general, computationally challenging. We focus on problems where resources enable agents to execute actions in stochastic environments, modeled as Markov decision processes (MDPs), such that the value of a resource bundle is defined as the expected value of the optimal MDP policy realizable given these resources. We present an algorithm that simultaneously solves the resource-allocation and the policy-optimization problems. This allows us to avoid explicitly representing utilities over exponentially many resource bundles, leading to drastic (often exponential) reductions in computational complexity. We then use this algorithm in the context of self-interested agents to design a combinatorial auction for allocating resources. We empirically demonstrate the effectiveness of our approach by showing that it can, in minutes, optimally solve problems for which a straightforward combinatorial resource-allocation technique would require the agents to enumerate up to 2^100 resource bundles and the auctioneer to solve an NP-complete problem with an input of that size.

We consider interactive tools that help users search for their most preferred item in a large collection of options. In particular, we examine example-critiquing, a technique for enabling users to incrementally construct preference models by critiquing example options that are presented to them. We present novel techniques for improving the example-critiquing technology by adding suggestions to its displayed options. Such suggestions are calculated based on an analysis of users' current preference model and their potential hidden preferences. We evaluate the performance of our model-based suggestion techniques with both synthetic and real users. Results show that such suggestions are highly attractive to users and can stimulate them to express more preferences to improve the chance of identifying their most preferred item by up to 78%.

Still, using symbolic representations of the building blocks of thinking, researchers in AI labs can execute programs that perform the sort of "symbol collision" that produces high-level thinking. The result enable the development of automated software of colliding beams of gold traveling near the technologies for assisting humans in these speed of light ("minibangs") allows physicists tasks. As with cosmology, the results of artificially to observe the liberation of quarks and gluons created intelligence validate theories of from protons and neutrons, revealing conditions intelligence of the natural kind. Theoretical unfolds in a set of often entertaining essays by and experimental breakthroughs since the philosophers, cognitive scientists, and computer 1970s, as well as technological advances in the scientists. The topics related to this theme art of colliding and detecting particles, have explored here are quite diverse and are typically made it possible to observe a new "energy frontier," presented in an informal, discursive style.

(MDAI 2007).

Artificial intelligence (AI), on the twenty-fifth anniversary of its naming, is a "kid, finally grown up." In this letter to his field, Feigenbaum recounts AI's stumbles and successes, its growing pains and maturation, to a place of preeminence among the sciences; standing with molecular biology, particle physics, and cosmology as owners of the best questions of science.