Search

Humans engaged in collaborative activities are naturally able to convey their intentions to teammates through multi-modal communication, which is made up of explicit and implicit cues. Similarly, a more natural form of human-robot collaboration may be achieved by enabling robots to convey their intentions to human teammates via multiple communication channels. In this paper, we postulate that a better communication may take place should collaborative robots be able to anticipate their movements to human teammates in an intuitive way. In order to support such a claim, we propose a robot system's architecture through which robots can communicate planned motions to human teammates leveraging a Mixed Reality interface powered by modern head-mounted displays. Specifically, the robot's hologram, which is superimposed to the real robot in the human teammate's point of view, shows the robot's future movements, allowing the human to understand them in advance, and possibly react to them in an appropriate way. We conduct a preliminary user study to evaluate the effectiveness of the proposed anticipatory visualization during a complex collaborative task. The experimental results suggest that an improved and more natural collaboration can be achieved by employing this anticipatory communication mode.

This paper investigates the policy optimization paradigm, where a learning model is trained to find the solution of complex Markov decision problems, as a tool to address the berth allocation problem in multimodal terminals. To this purpose, we drop the typical formulation of the latter as a mixed-integer static scheduling one, and we model it instead as an evolving scenario in which berths are assigned to ships according to a parameterized policy function that drives the temporal evolution of the environment. We adopt a cross-entropy optimization scheme to optimize the policy parameters, which is a simple and highly parallelizable gradient-free technique. As compared to the static mixed-integer formulation, the proposed approach relies on a much lighter optimization problem in the continuous space of the policy parameters, thus making it feasible to replan in real time when needed. Furthermore, the generality of the policy optimization approach allows to take into account any performance metric and specific feature of the scenario straightforwardly, without the need to devise ad hoc heuristics. Simulation tests showcase the good performance of the policy approach under various conditions.

In this paper we investigate deep learning architectures combined with low-discrepancy sampling as surrogate models. The aim is to provide a quick estimate of the outcome of an expensive process when many evaluations are needed, e.g., for optimization purposes. The simulation of urban traffic, routinely employed for the design and performance evaluation of policies in urban mobility scenarios, is an example of procedure characterized by very complex dynamics, large dimensionality and intensive computational requirements. In this context, the need for a surrogate model arises in many forms, e.g., for origin-destination demand calibration, traffic light optimization, strategic planning. In order to cope with this complexity and large dimensionality we take advantage of the excellent approximating capabilities of deep neural networks as learning models. Then, we employ low-discrepancy sequences as sampling designs for the simulation runs required to create the training set for the deep surrogate model. This kind of sampling guarantees deterministically a good covering of the input space, and is able to exploit possible regularities of the simulation outcome. Extensive experimental tests are presented involving the popular SUMO microsimulator, showing the advantages of the proposed surrogate modeling solution under various performance measures.

Energy demands of modern datacentres are an immense concern. An intuitive solution is to turn servers off to incur less costs. However, the control problem of when to turn a specific server off, and when to then turn that server back on, is far from trivial. As such, many different authors have modeled this problem as an $M/M/C$ queue where each server can be turned on, with an exponentially distributed setup time, or turned off instantaneously. We analyse this well-established model under the asymptotic regime where the number of servers approaches infinity while the load per server remains fixed and show that not only are many of the control policies in the literature equivalent under this regime, but they are also optimal under any cost function which is non-decreasing in the expected energy cost and response time.

Energy consumption of today's datacenters is a constant concern from the standpoints of monetary and environmental costs. We model a datacenter as a queueing system, where each server can be switched on or off, with the time to switch a server on being nonnegligible. Previously derived structural properties of the optimal policy allow us to intelligently select policies to analyse further. Using the recursive renewal reward technique, we offer an exact analysis of these policies alongside offering insights, observations, and implications for how these systems behave. In particular, we provide insight into the question of the number of servers that should remain on at all times under a general cost function.

Over the past few years, energy provisioning in server farms and data-centres has become an active area of research. As such, many models have been proposed where an individual server has setup times and can switch between two different energy states (on and off). To make such models tractable, assumptions are usually made on the type of policies the system can implement. However, it is often not known if such assumptions allow for the model to capture the optimal policy, or if such a model will be strictly suboptimal. In this work we model such systems using Markov Decision Processes (MDPs) and derive several structural properties which (partially) describe the optimal policy. These properties reduce the set of feasible policies significantly, allowing one to describe the optimal policy by a set of thresholds which have considerable structure. In addition to the analysis, we discuss the current literature in the context of our results.

This work addresses the problem of learning an unknown function from data when local models are employed. In particular, kernel smoothing models are considered, which use kernels in a straightforward fashion by modeling the output as a weighted average of values observed in a neighborhood of the input. Such models are a popular alternative to other kernel paradigms, such as support vector machines (SVM), due to their very light computational burden. The purpose of this work is to prove that a smart deterministic selection of the observation points can be advantageous with respect to input data coming from a pure random sampling. Apart from the theoretical interest, this has a practical implication in all the cases in which one can control the generation of the input samples (e.g., in applications from robotics, dynamic programming, optimization, mechanics, etc.) To this purpose, lattice point sets (LPSs), a special kind of sampling schemes commonly employed for efficient numerical integration, are investigated. It is proved that building local kernel smoothers using LPSs guarantees universal approximation property with better rates with respect to i.i.d. sampling. Then, a rule for automatic kernel width selection, making the computational burden of building the model negligible, is introduced to show how the regular structure of the lattice can lead to practical advantages. Simulation results are also provided to test in practice the performance of the proposed methods.

When the problem of learning from data is solved through a regression tree estimator, the quality of the available observations is an important issue, since it influences directly the accuracy of the resulting model. It becomes particuarly relevant when there is freedom to sample the input space arbitrarily to build the tree model or, alternatively, when we need to select a subsample to train the tree estimator on a computationally feasible input set, or to evaluate the goodness of the estimation on a test set. Here the accuracy of estimation based on regression trees is analyzed from the point of view of geometric properties of the available input data. In particular, the concept of F-discrepancy, a quantity that measures how well a set of points represents the distribution underlying the input generation process, is applied to derive conditions for convergence to the optimal piecewise-constant estimator for the unknown function we want to learn. The analysis has a constructive nature, allowing to select in practice good input sets for the problem at hand, as shown in a simulation example involving a real data set.

In this work local linear regression models are introduced and analyzed in the context of empirical risk minimization (ERM) for function learning. This kind of models can be seen as a more sophisticated version of classic kernel smoothing models, based on the principle of local estimation. In particular, we analyze the conditions under which consistency of the ERM procedure is guaranteed, pointing out assumptions on the way the input space is sampled to obtain the observation data. This allows to extend the tractation to the case where the choice of the training set is part of the learning process. To this purpose, a choice of the observation points based on low-discrepancy sequences, a family of sampling methods commonly employed for efficient numerical integration, is analyzed. Simulation results involving two different examples of function learning are provided.

This paper analyzes quasi-random sampling techniques for approximate dynamic programming. Specifically, low-discrepancy sequences and lattice point sets are investigated and compared as efficient schemes for uniform sampling of the state space in high-dimensional settings. The convergence analysis of the approximate solution is provided basing on geometric properties of the two discretization methods. It is also shown that such schemes are able to take advantage of regularities of the value functions, possibly through suitable transformations of the state vector. Simulation results concerning optimal management of a water reservoirs system and inventory control are presented to show the effectiveness of the considered techniques with respect to pure-random sampling.

As energy costs and energy used by server farms increase, so does the desire to implement energy-aware policies. Although under some metrics, optimal policies for single as well as multiple server systems are known, a number of metrics remain without sufficient knowledge of corresponding optimal policies. We describe and analyse a model to determine an optimal policy for on/off single server systems under a broad range of cost functions that are based on expected response time, expected energy costs, and expected wear and tear costs. We leverage this model in the problem of routing jobs to one of two servers to show a range of non-trivial optimal routing probabilities and server configurations when energy concerns are a factor.

The design of observers and controllers for a class of continuous-time, nonlinear dynamic systems with Lipschitz nonlinearities is addressed. Observers and controllers that depend on a linear gain and a parameterized function implemented by a feedforward neural network are considered. The gain and the weights of the neural network are optimized in such way to ensure the convergence of the estimation error for the observer and the stability of the closed-loop system for the controller, respectively. This is achieved by constraining the derivative of a quadratic Lyapunov function to be negative definite on a grid of points, penalizing the constraints that are not satisfied. It is shown that suitable sampling techniques such as low-discrepancy sequences, commonly employed in quasi-Monte Carlo methods for high-dimensional integration, allow one to reduce the computational burden required to optimize the network parameters. Simulations results are presented to illustrate the effectiveness of the method. © 2006 IEEE.

Summary form only given. UML (Unified Modelling Language) is a universal language for modelling of object-oriented applications. UML can build models of different kinds of domains. However, no paradigm is enough to clearly model all views of all possible domains. For this reason, there are several metamodel elements to extend UML and facilitate the modelling of specific domains. We analyze three kinds of existing extensions to the UML metamodel and propose another extension, evolutionary stereotype. The abstract syntax, well-formedness rules and semantics are defined. A case study is written to show how the dynamic semantics is specified. These stereotypes allow integrating not only the possibility to specify restrictions to the metamodel, in particular using OCL (Object Constraint Language), but also to aggregate new semantic definitions. The UML metamodel is organized in logic packages. One of them is the extension mechanisms package where the extensions to UML are specified. The evolutionary stereotypes are defined as a new metaclass of the extension mechanisms package. This new metaclass has associations that relate this metaclass with existent metaclasses of the OMG (Object Management Group) metamodel.

A standard EG&G VMCM (vector measuring current meter) has been modified to acquire data from external sensors. The external sensors are mounted on a compact backpack which attaches to the VMCM's cage. Some simple changes to the VMCM's A/D (analog/digital) board opened up eight analog voltage channels. Changes to the VMCM's firmware were made to allow control over the period of time between toggling external power on and the A/D sampling cycle. Some sensors with low power consumption were powered continuously and their signals passed through a low-pass filter to match time constants and minimize aliasing. Sixteen VMCMs were modified and three deployments of these MVMSs (multivariable moored sensors) were made in 1987 with durations of three to four months. >