We propose Action-Reaction Learning as an approach for analyzing and synthesizing human behaviour. This paradigm uncovers causal mappings between past and future events or between an action and its reaction by observing time sequences. We apply this method to analyze human interaction and to subsequently synthesize human behaviour. Using a time series of perceptual measurements, a system automatically uncovers correlations between past gestures from one human participant (an action) and a subsequent gesture(areaction) from another participant. A probabilistic model is trainedfrom data of the human interaction using a novel estimation technique, Conditional Expectation Maximization (CEM). The estimation uses general bounding and maximization to monotonically find the maximum conditional likelihood solution. The learning system drives a graphical interactive character which probabilistically predicts a likely response to a user's behaviour and performs it interactively. Thus, after analyzing human interaction in a pair of participants, the system is able to replace one of them and interact with a single remaining user. 1

Problem solvers often have multiple operators available to them but must select just one to apply. We present three experiments that demonstrate that solvers use at least two sources of information to make operator selections in the building sticks task (BST): information from their past history of using the operators and information from the current context of the problem. Specifically, problem solvers are more likely to use an operator the more successful it has been in the past and the closer it takes the current state to the goal state. These two effects, respectively, represent the learning and performance processes that influence solvers ’ operator selections. A computational model of BST problem solving, developed within the ACT-R theory (Anderson, 1993), provides the unifying framework in which both types of processes can be integrated to predict solvers ’ selection tendencies. � 1996 Academic Press, Inc. Most problems can be approached in multiple ways but solved by only a few. Problem solving can be viewed, then, as finding one of the few paths that leads from a problem’s initial state to its goal state through some space of possible intermediate states (Newell & Simon, 1972). In this framework,

In unidimensional absolute identification tasks, participants identify stimuli that vary along a single dimension. Performance is surprisingly poor compared with discrimination of the same stimuli. Existing models assume that identification is achieved using long-term representations of absolute magnitudes. The authors propose an alternative relative judgment model (RJM) in which the elemental perceptual units are representations of the differences between current and previous stimuli. These differences are used, together with the previous feedback, to respond. Without using long-term representations of absolute magnitudes, the RJM accounts for (a) information transmission limits, (b) bowed serial position effects, and (c) sequential effects, where responses are biased toward immediately preceding stimuli but away from more distant stimuli (assimilation and contrast).

We propose Action-Reaction Learning as an approach for analyzing and synthesizing human behaviour. This paradigm uncovers causal mappings between past and future events or between an action and its reaction by observing time sequences. We apply this methodto analyze human interaction and to subsequently synthesize human behaviour. Using a time series of perceptual measurements, a system automatically uncovers a mapping between gestures from one human participant (an action) and a subsequent gesture(areaction) from another participant. Aprobabilistic model is trained from data of the human interaction using a novel estimation technique, Conditional Expectation Maximization (CEM). The system drives a graphical interactive character which probabilistically predicts the most likely response to the user's behaviour and performs it interactively. Thus, after analyzing human interaction in a pair of participants, the system is able to replaceoneof them and interact with a single remaining user.

Categorization research typically assumes that the cognitive system has access to a (more or less noisy) representation of the absolute magnitudes of the properties of stimuli and that this information is used in reaching a categorization decision. However, research on identification of simple perceptual stimuli suggests that people have very poor representations of absolute magnitude information and that judgments about absolute magnitude are strongly influenced by preceding material. The experiments presented here investigate such sequence effects in categorization tasks. Strong sequence effects were found. Classification of a borderline stimulus was more accurate when preceded by a distant member of

Two basic approaches to explaining the nature of the mind are the rational and mechanistic approaches. Rational analyses attempt to characterize the environment and the behavioral outcomes that humans seek to optimize, whereas mechanistic models attempt to simulate human behavior using processes and representations analogous to those used by humans. We compared these approaches on their accounts of how humans learn the variability of categories. The mechanistic model departs in subtle ways from rational principles. In particular, the mechanistic model incrementally updates its estimates of category means and variances through error-driven learning, based on discrepancies between new category members and the current representation of each

I propose Action-Reaction Learning as an approach for analyzing and synthesizing human behaviour. This paradigm uncovers causal mappings between past and future events or between an action and its reaction by observing time sequences. I apply this method to analyze human interaction and to subsequently synthesize human behaviour. Using a time series of perceptual measurements, a system automatically uncovers a mapping between past gestures from one human participant (an action) and a subsequent gesture (a reaction) from another participant. A probabilistic model is trained from data of the human interaction using a novel estimation technique, Conditional Expectation Maximization (CEM). The estimation uses general bounding and maximizationtofindthemaximum conditional likelihood solution. The learning system drives a graphical interactivecharacter which probabilistically predicts the most likely response to a user's behaviour and performs it interactively.Thus, after analyzing human interaction in a pair of participants, the system is able to replace one of them and interact with a single remaining user. Thesis Supervisor: Alex Pentland Title: Academic Head and Toshiba Professor of Media Arts and Sciences, MIT Media Lab This work was supported in part by British Telecom and Texas Instruments. Action-Reaction Learning: Analysis and Synthesis of Human Behaviour by Tony Jebara The following people served as readers for this thesis: Reader: Bruce M. Blumberg Asahi Broadcasting Corporation Career Development Assistant Professor of Media Arts and Sciences MIT Media Laboratory Reader: Aaron Bobick Assistant Professor of Computational Vision MIT Media Laboratory 4 Acknowledgments I extend warm thanks to my advisor, Professor Alex Pentland for having given me to opport...

Multiple cue probability learning studies have typically focused on stationary environments. We present three experiments investigating learning in changing environments. A fine-grained analysis of the learning dynamics shows that participants were responsive to both abrupt and gradual changes in cue-outcome relations. We found no evidence that participants adapted to these types of change in qualitatively different ways. Also, in contrast to earlier claims that these tasks are learned implicitly, participants showed good insight into what they learned. By fitting formal learning models, we investigated whether participants learned global functional relationships or made localized predictions from similar experienced exemplars. Both a local (the Associative Learning Model) and a global learning model (the novel Bayesian Linear Filter) fitted the data of the first two experiments. However, the results of Experiment 3, which was specifically designed to discriminate between local and global learning models, provided more support for global learning models. Finally, we present a novel model to account for the cue competition effects found in previous research and displayed by some of our participants.