The ``comma sequence'' starts with 1 and is defined by the property that if k and k' are consecutive terms, the two-digit number formed from the last digit of k and the first digit of k' is equal to the difference k'-k. If there is more than one such k', choose the smallest, but if there is no such k' the sequence terminates. The sequence begins 1, 12, 35, 94, 135, ... and, surprisingly, ends at term 2137453, which is 99999945. The paper analyzes the sequence and its generalizations to other starting values and other bases. A slight change in the rules allows infinitely long comma sequences to exist.

We suggest that as people move to construe robots as social agents, interact with them, and treat them as capable of social ties, they might develop (close) relationships with them. We then ask what kind of relationships can people form with bots, what functions can bots fulfill, and what are the societal and moral implications of such relationships.

Machine learning holds the potential to be a powerful tool to aid in designing catalytic and sustainable chemical systems. However, it is important for experimental researchers to understand the capabilities of different machine learning models when trained on experimental data. In this work, we trained three different machine learning algorithms (decision tree, random forest, and multilayer perceptron) with a hand-curated dataset of 127 reaction conditions for electrocatalytic CO2 reduction on heterogeneous catalysts in aqueous electrolytes. The input to the machine learning models were the experimental conditions, and we posed four separate outputs to each of these machine learning algorithms: (1) if the number of proton-coupled electron transfer events was greater than two, (2) if carbon-carbon coupling occurred, (3) if ethylene was the major product, and (4) major product prediction. We observed that with a dataset of this size, all three machine learning models could achieve accuracies between 0.7 and 0.8 for the three binary classification problems (1, 2, and 3). Also, the shallow learning decision tree and random forest models performed equal to or better than the deep learning multilayer perceptron models. In the multiclass classification problem (i.e., predicting the product) the accuracy for all models decreased, with the random forest model producing the highest accuracy of 0.6. Analysis of the models showed that machine learning can independently arrive at conclusions that are well-known in the literature, e.g., that Cu is an important catalyst for producing high-carbon content products, and discern more-complicated patterns, with respect to feature importance.

Lack of trust is one of the main obstacles standing in the way of taking full advantage of the benefits artificial intelligence (AI) has to offer. Most research on trust in AI focuses on cognitive ways to boost trust. Here, instead, we focus on boosting trust in AI via affective means. Specifically, we tested and found associations between one's attachment style-an individual difference representing the way people feel, think, and behave in relationships-and trust in AI. In Study 1 we found that attachment anxiety predicted less trust. In Study 2, we found that enhancing attachment anxiety reduced trust, whereas enhancing attachment security increased trust in AI. In Study 3, we found that exposure to attachment security cues (but not positive affect cues) resulted in increased trust as compared with exposure to neutral cues. Overall, our findings demonstrate an association between attachment security and trust in AI, and support the ability to increase trust in AI via attachment security priming.

In active sensing, sensing actions are typically chosen to minimize the uncertainty of the state according to some information-theoretic measure such as entropy, conditional entropy, mutual information, etc. This is reasonable for applications where the goal is to obtain information. However, when the information about the state is used to perform a task, minimizing state uncertainty may not lead to sensing actions that provide the information that is most useful to the task. This is because the uncertainty in some subspace of the state space could have more impact on the performance of the task than others, and this dependence can vary at different stages of the task. One way to combine task, uncertainty, and sensing, is to model the problem as a sequential decision making problem under uncertainty. Unfortunately, the solutions to these problems are computationally expensive. This paper presents a new task-oriented active sensing scheme, where the task is taken into account in sensing action selection by choosing sensing actions that minimize the uncertainty in future task-related actions instead of state uncertainty. The proposed method is validated via simulations.

Functional Electrical Stimulation (FES) employs neuroprostheses to apply electrical current to the nerves and muscles of individuals paralyzed by spinal cord injury to restore voluntary movement. Neuroprosthesis controllers calculate stimulation patterns to produce desired actions. To date, no existing controller is able to efficiently adapt its control strategy to the wide range of possible physiological arm characteristics, reaching movements, and user preferences that vary over time. Reinforcement learning (RL) is a control strategy that can incorporate human reward signals as inputs to allow human users to shape controller behavior. In this paper, ten neurologically intact human participants assigned subjective numerical rewards to train RL controllers, evaluating animations of goal-oriented reaching tasks performed using a planar musculoskeletal human arm simulation. The RL controller learning achieved using human trainers was compared with learning accomplished using human-like rewards generated by an algorithm; metrics included success at reaching the specified target; time required to reach the target; and target overshoot. Both sets of controllers learned efficiently and with minimal differences, significantly outperforming standard controllers. Reward positivity and consistency were found to be unrelated to learning success. These results suggest that human rewards can be used effectively to train RL-based FES controllers.

High-level spinal cord injury (SCI) in humans causes paralysis below the neck. Functional electrical stimulation (FES) technology applies electrical current to nerves and muscles to restore movement, and controllers for upper extremity FES neuroprostheses calculate stimulation patterns to produce desired arm movement. However, currently available FES controllers have yet to restore natural movements. Reinforcement learning (RL) is a reward-driven control technique; it can employ user-generated rewards, and human preferences can be used in training. To test this concept with FES, we conducted simulation experiments using computer-generated "pseudohuman" rewards. Rewards with varying properties were used with an actor-critic RL controller for a planar two-degree-of-freedom biomechanical human arm model performing reaching movements. Results demonstrate that sparse, delayed pseudo-human rewards permit stable and effective RL controller learning. The frequency of reward is proportional to learning success, and human-scale sparse rewards permit greater learning than exclusively automated rewards. Diversity of training task sets did not affect learning. Longterm stability of trained controllers was observed. Using human-generated rewards to train RL controllers for upper-extremity FES systems may be useful. Our findings represent progress toward achieving human-machine teaming in control of upper-extremity FES systems for more natural arm movements based on human user preferences and RL algorithm learning capabilities.

We consider a linear consensus system with n agents that can switch between r different connectivity patterns. A natural question is which switching law yields the best (or worst) possible speed of convergence to consensus? We formulate this question in a rigorous manner by relaxing the switched system into a bilinear consensus control system, with the control playing the role of the switching law. A best (or worst) possible switching law then corresponds to an optimal control. We derive a necessary condition for optimality, stated in the form of a maximum principle (MP). Our approach, combined with suitable algorithms for numerically solving optimal control problems, may be used to obtain explicit lower and upper bounds on the achievable rate of convergence to consensus. We also show that the system will converge to consensus for any switching law if and only if a certain (n-1) dimensional linear switched system converges to the origin for any switching law. For the case n=3 and r=2, this yields a necessary and sufficient condition for convergence to consensus that admits a simple graph-theoretic interpretation.

We present a versatile application for cyber-physical systems (CPS), called the Cloud Conveyors System (CCS). This system comprises a collection of mobile conveyor units with simple periodic behavior; the units move back and forth along fixed tracks. The system-level objective is to transport entities from some input port to an output port when each entity has its own target output port, deadline, and end-to-end QoS constraints. Entities ride on the mobile units to physically move from one location to another. Entities may transfer instantaneously between two units — or when the unit is at an input or an output. We refer to these transfers as cyber transfers because they involve decision-making and the entities do not have to transfer at every possible opportunity. We view the transport of each entity in CCS as a CPS-Task that evolves both in space and in time; more precisely, a CPS-Task is an alternating sequence of cyber transfers and physical moves that starts at an input and ends at the output of the entity. This novel model for a CPS-Task allows one to explore solutions to some of the principal CPS challenges namely, Composition, Control Strategies, Computational Abstractions, Model-driven Engineering, and Verification & Validation. Further, this abstract and well-defined problem is useful in CPS Education and Training because it has a rich structure with intertwined cyber and physical dynamics; also, the scale and complexity of the problem can be increased by adding more units or changing the configuration of the system without increasing the implementation burden, which is critical to validating CPS solution techniques on physical testbeds.

Cockroach shelter-seeking strategy may initially look like an undirected random search, but we show that they are attracted to darkened shelters. They arrive at a shelter in about half the time control cockroaches take to reach the same location with no shelter present. We were able to identify six statistically significant trends from the behavior of 134 cockroaches in 1-min naïve walking trials with four different shelter configurations. By combining these trends into a model, we built a stochastic algorithm that significantly biases a simulated agent toward a target location. We call this model RAMBLER (Randomized Algorithm Mimicking Biased Lone Exploration in Roaches). RAMBLER could be adapted for a mobile robot equipped with an onboard camera and antenna-like contact sensors.

Author Index Ahmadi, Seyed Alireza 179 Ames, Aaron D. 219 Amoozadeh, Mani 179 Annaswamy, Anuradha M. 339 Araujo, José 179 Baras, John S. 23 Basar, Tamer 301 Bezzo, Nicola 197 Bolognani, Saverio 259 Branicky, Michael S. 43 Bushnell, Linda 161, 301 Cassandras, Christos G. 281 Cavraro, Guido 259 Chakraborty, Samarjit 339 Chapman, Airlie 143 Clark, Andrew 161 Cortés, Jorge 317 Cvijic, Sanja 241 Elia, Nicola 357 Goswami, Dip 339 Hartman, Matthew 3 Hespanha, João P. 85 Ilic, Marija 241 Johansson, Karl Henrik 123, 179 Jones, Malachi 65 Kemmerer, Richard A. 85 Kotsalis, Georgios 65 Lee, Insup 197 Lee, Phillip 161 Lin, Xuchao 281 Ma, Xu 357 Malik, Waseem A. 101 Martins, Nuno C. 101 Mesbahi, Mehran 143 Nowzari, Cameron 317 Pajic, Miroslav 197 Pappas, George J. 197 Poovendran, Radha 161 Powell, Matthew 219 Sandberg, Henrik 123, 179 Sastry, PS 43 Sastry, Shivakumar …

When creating a genetic algorithm one must decide how the phenotype is encoded in the genotype. This decision can change the "evolvability" of the system, that is, how quickly high fitness solutions are found. A change in encoding can cause high fitness solutions to be found more quickly or result in no high fitness solutions found at all.Direct and indirect encodings are two general types of encodings. Direct encodings are characterized by one genotype parameter mapping to one phenotype parameter. Indirect encodings are characterized by a one-to-many genotype to phenotype parameter mapping.Researchers have often demonstrated that indirect encodings have higher evolvability than direct encodings on a variety of problems. The indirect encodings were hypothesized to be more evolvable for two main reasons, both a consequence of the one-to-many property of the encoding. First, when parts of the genotype are reused, a given phenotype can often be described with a smaller genotype. A smaller genotype space should be faster to search than a larger genotype space. Second, the one-to-many property often leads to regularity in the phenotype produced by indirect encodings. This may be useful in exploiting regularities in the problems used for demonstrating that indirect encodings have higher evolvabilities than direct encodings.This thesis explores whether reusing parts of the genotype increases evolvability more than size of the genotype for an example system. It is found that genotype reuse is more important than genotype size.The thesis also explores whether an indirect encoding can increase evolvability and be selected for on a problem on which high fitness can be achieved with irregular solutions. It is found that it is very difficult to construct a problem which does not contain some form of regularity which can be exploited by the indirect encoding to increase evolvability. A problem which appears not to have regularity is found to have non-structural regularitiy. Based on this understanding, the problem is modified and the indirect encoding is found to have no or a slightly negative impact on evolvability on the modified problem.

Cockroach shelter-seeking strategy may look like an undirected random search, but we show that they are attracted to darkened shelters, arriving at a shelter in about half the time it would otherwise take. We were able to identify four statistically significant trends from the behavior of 134 cockroaches in one-minute naïve walking trials with four different arena configurations. By combining these trends into a model, we arrive at an algorithm that significantly directs a simulated agent to a location. This algorithm was then adapted and tested on a small mobile robot equipped with an onboard camera and antenna-like contact sensors.

Advanced Manufacturing systems, such as reconfigurable conveyor systems, are critical to several economies and societies. They enable the rapid reconfiguration of the system at the level of individual units or subsystems to cope with the needs of emerging markets and applications. However, because of the immense scale of these systems, several components tend to fail on a regular basis. There is an urgent need for autonomous fault tolerance techniques that will ensure high availability of these critical cyber physical systems. In this paper, we present preliminary results from our design of a highly available reconfigurable conveyor system. At the core of our approach are ideas for fault tolerance derived from TCP sliding window flow control and congestion control approaches.

The evolution of sampling-based planning has introduced a number of novel algorithms and modifications that allow for various adaptations to the local environment. This paper presents a revision to the Path-length Annexed Random Tree (PART) that allocates and regulates local exploration in an adaptive manner. The improved algorithm eliminates the need to choose and tune an additional threshold parameter while providing denser coverage in potentially complex regions and a natural means of connection for either unidirectional or bidirectional planning. The robustness of this algorithm and the performance implications of the threshold parameter are demonstrated in 2D and SE(3) experiments.

Robotic motion planning, which concerns the computation of paths and controls that drive an autonomous agent from one configuration to another, is quickly becoming a vitally important field of research as its applications diversify and become increasingly public. Many algorithms have been proposed to deal with this central problem; sampling-based approaches like the Rapidly-exploring Random Tree (RRT) and Probabilistic Roadmap Method (PRM) planners are among the most successful. Still, these algorithms are not fully understood and suffer from pathologically poorly-performing instances resulting from the contributions of random sampling and qualitative obstacle features like narrow passages. The large means and variances that result from these issues continue to motivate the development of new algorithms and adaptations to increase consistency and to allow more difficult problems to be solved.This research examines these performance issues with a focus on the Rapidly-exploring Random Tree (RRT) planner. Fundamental analysis establishes that the interaction of its Voronoi bias with particular obstacle features can compromise its efficacy and illustrates the types of distributions on its performance that result. It further provides guidance on the types of problems amenable to solutions by the algorithm and on the use of its alternative EXTEND and CONNECT heuristics and step size parameter. Observations from this analysis prompt an investigation of the use of restart strategies to manage issues of both scaling in computation and exploratory missteps. In turn, their impact provides a foundation for the introduction of a novel algorithm, the Path-length Annexed Random Tree (PART) planner, that directs its exploration on a local basis. This algorithm and its environment-adaptive successor, the Adaptive PART (APART) planner, demonstrate competitive performance on instructive examples and dramatic improvements on difficult benchmarks, while also supplementing their utility with the output of a connected roadmap.