Active learning of continuous-time Bayesian networks through interventions*

We consider the problem of learning structures and parameters of continuous-time Bayesian networks (CTBNs) from time-course data under minimal experimental resources. In practice, the cost of generating experimental data poses a bottleneck, especially in the natural and social sciences. A popular approach to overcome this is Bayesian optimal experimental design (BOED). However, BOED becomes infeasible in high-dimensional settings, as it involves integration over all possible experimental outcomes. We propose a novel criterion for experimental design based on a variational approximation of the expected information gain. We show that for CTBNs, a semi-analytical expression for this criterion can be calculated for structure and parameter learning. By doing so, we can replace sampling over experimental outcomes by solving the CTBNs master-equation, for which scalable approximations exist. This alleviates the computational burden of integrating over possible experimental outcomes in high-dimensions. We employ this framework in order to recommend interventional sequences. In this context, we extend the CTBN model to conditional CTBNs in order to incorporate interventions. We demonstrate the performance of our criterion on synthetic and real-world data.

A Sequential Inspection Procedure for Fault Detection in Multistage Manufacturing Processes

Fault diagnosis in multistage manufacturing processes (MMPs) is a challenging task where most of the research presented in the literature considers a predefined inspection scheme to identify the sources of variation and make the process diagnosable. In this paper, a sequential inspection procedure to detect the process fault based on a sequential testing algorithm and a minimum monitoring system is proposed. After the monitoring system detects that the process is out of statistical control, the features to be inspected (end of line or in process measurements) are defined sequentially according to the expected information gain of each potential inspection measurement. A case study is analyzed to prove the benefits of this approach with respect to a predefined inspection scheme and a randomized sequential inspection considering both the use and non-use of fault probabilities from historical maintenance data.

Exploration and Coordination of Complementary Multirobot Teams in a Hunter-and-Gatherer Scenario

The hunter-and-gatherer approach copes with the problem of dynamic multirobot task allocation, where tasks are unknowingly distributed over an environment. This approach employs two complementary teams of agents: one agile in exploring (hunters) and another dexterous in completing (gatherers) the tasks. Although this approach has been studied from the task planning point of view in our previous works, the multirobot exploration and coordination aspects of the problem remain uninvestigated. This paper proposes a multirobot exploration algorithm for hunters based on innovative notions of “expected information gain” to minimize the collective cost of task accomplishments in a distributed manner. Besides, we present a coordination solution between hunters and gatherers by integrating the novel notion of profit margins into the concept of expected information gain. Statistical analysis of extensive simulation results confirms the efficacy of the proposed algorithms compared in different environments with varying levels of obstacle complexities. We also demonstrate that the lack of effective coordination between hunters and gatherers significantly distorts the total effectiveness of the planning, especially in environments containing dense obstacles and confined corridors. Finally, it is statistically proven that the overall workload is distributed equally for each type of agent which ensures that the proposed solution is not biased to a particular agent and all agents behave analogously under similar characteristics.

A Complete Method for Assessing the Effectiveness of Eyewitness Identification Procedures: Expected Information Gain

We present a method for measuring the efficacy of eyewitness identification procedures by applying fundamental principles of information theory. The resulting measure evaluates the Expected Information Gain (EIG) for an identification attempt, a single value that summarizes an identification procedure’s overall potential for reducing uncertainty about guilt or innocence across all possible witness responses. In a series of demonstrations, we show that EIG often disagrees with existing measures (e.g., diagnosticity ratios or area under the ROC) about the relative effectiveness of different identification procedures. Each demonstration is designed to highlight key distinctions between existing measures and EIG. An overarching theme is that EIG provides a complete measure of evidentiary value, in the sense that it factors in all aspects of identification performance. Collectively, these demonstrations show that EIG has substantial potential to inspire new discoveries in eyewitness research and provide a new perspective on policy recommendations for the use of identifications in real investigations.

In Pursuit of Knowledge: Preschoolers Expect Agents to Weigh Information Gain and Information Cost When Deciding Whether to Explore

When deciding whether to explore, agents must consider both their need for information and its cost. Do children recognize that exploration reflects a trade-off between action costs and expected information gain, inferring epistemic states accordingly? In two experiments, 4- and 5-year-olds (N=144; of diverse race and ethnicity) judge that an agent who refuses to obtain low-cost information must have already known it, and an agent who incurs a greater cost to gain information must have a greater epistemic desire. Two control studies suggest that these findings cannot be explained by low-level associations between competence and knowledge. Our results suggest that preschoolers’ Theory of Mind includes expectations about how costs interact with epistemic desires and states to produce exploratory action.

A Complete Method for Assessing the Effectiveness of Eyewitness Identification Procedures: Expected Information Gain

We present a method for measuring the efficacy of eyewitness identification procedures by applying fundamental principles of information theory. The resulting measure evaluates the Expected Information Gain (EIG) for an identification attempt, a single value that summarizes an identification procedure’s overall potential for reducing uncertainty about guilt or innocence across all possible witness responses. In a series of theoretical demonstrations, we show that EIG often disagrees with existing measures (e.g., diagnosticity ratios or area under the ROC) about the relative effectiveness of different identification procedures. Each demonstration is designed to highlight “blind spots” of the existing measures as a contrast to EIG, which considers every factor relevant to a procedure’s potential for decreasing uncertainty about guilt or innocence. Collectively, these demonstrations show that EIG has substantial potential to inspire new discoveries in eyewitness research. For research designed to identify procedures that will be most effective in criminal investigations, EIG supersedes all other measures, on both theoretical and practical grounds.

Project Management Monitoring Based on Expected Duration Entropy

Projects are rarely executed exactly as planned. Often, the actual duration of a project’s activities differ from the planned duration, resulting in costs stemming from the inaccurate estimation of the activity’s completion date. While monitoring a project at various inspection points is pricy, it can lead to a better estimation of the project completion time, hence saving costs. Nonetheless, identifying the optimal inspection points is a difficult task, as it requires evaluating a large number of the project’s path options, even for small-scale projects. This paper proposes an analytical method for identifying the optimal project inspection points by using information theory measures. We search for monitoring (inspection) points that can maximize the information about the project’s estimated duration or completion time. The proposed methodology is based on a simulation-optimization scheme using a Monte Carlo engine that simulates potential activities’ durations. An exhaustive search is performed of all possible monitoring points to find those with the highest expected information gain on the project duration. The proposed algorithm’s complexity is little affected by the number of activities, and the algorithm can address large projects with hundreds or thousands of activities. Numerical experimentation and an analysis of various parameters are presented.

Exploratory play, rational action, and efficient search

Play is a universal behavior widely held to be critical for learning and development. Recent studies suggest children’s exploratory play is consistent with formal accounts of learning. This “play as rational exploration” view suggests that children’s play is sensitive to costs, rewards, and expected information gain. By contrast, here we suggest that a defining feature of human play is that children subvert normal utility functions in play, setting up problems where they incur needless costs to achieve arbitrary rewards. Across four studies, we show that 4-5-year-old children not only infer playful behavior from observed violations of rational action (Experiment 1), but themselves take on unnecessary costs and perform inefficient actions during play, despite understanding and valuing efficiency in non-playful, instrumental contexts (Experiment 2-4). We end with a discussion of the value of apparently utility-violating behavior and why it might serve learning in the long run.