Flagging hate speech on social media messages has important societal benefits. While large language models have become increasingly able to identify hate speech with high accuracy, they come with a significant computational cost. Thus, there is a need for simpler models that detect hateful or abusive content by classifying an embedding of the text. Such models perform very well for explicitly abusive content, but struggle with the classification of implicit hate. Furthermore, it has been found that the performance decreases significantly in cross-dataset experiments. In this paper, we propose Closest Positive Cluster (CPC) an auxiliary loss that increases the generalizability of embedding-based explicit and implicit hate classifiers in crossdataset scenarios. Through experiments spanning ten different hate speech datasets, we found that the CPC loss increased the model performance by 0.17 - 7.4% when added to the binary crossentropy loss during training. The experiments also investigated whether models trained on specific hate speech datasets generalize better to other datasets.

A common class of algorithms for informative path planning (IPP) follows boustrophedon ("as the ox turns") patterns, which aim to achieve uniform area coverage. However, IPP is often applied in scenarios where anomalies, such as plant diseases, pollution, or hurricane damage, appear in clusters. In such cases, prioritizing the exploration of anomalous regions over uniform coverage is beneficial. This work introduces a class of algorithms referred to as bounomōdes ("as the ox grazes"), which alternates between uniform boustrophedon sampling and targeted exploration of detected anomaly clusters. While uniform sampling can be designed using geometric principles, close exploration of clusters depends on the spatial distribution of anomalies and must be learned. In our implementation, the close exploration behavior is learned using deep reinforcement learning algorithms. Experimental evaluations demonstrate that the proposed approach outperforms several established baselines.

Gas distribution mapping (GDM) refers to the task of mapping the gas concentrations of an airborne chemical over a region of interest. A mobile robot equipped with a gas sensor can be used potentially autonomously to build such a distribution map. However, modern-day robots might not have enough battery power to cover the entire area of interest. Therefore, a group of n such collaborative mobile robots can be used for this purpose. The goal of the robots is to sample concentrations from a fraction of locations and infer the gas intensities in the rest of the area using a supervised machine learning technique, namely the Gaussian Process (GP). To this end, we propose a novel multi-robot gas distribution mapping framework, named GDM-Net++, which works in both 2D and 3D settings. Our proposed framework first divides the environment into n unique regions using Voronoi partitioning. Next, we employ a multi-agent deep Q-learning framework for the robots to learn a joint policy. As GP is a compute-intensive process, during testing, the learned policy is applied without re-training the GP model. The experiments are performed in simulation using Python on six types of Gaussian plumes to validate our proposed technique. Compared to two baselines – greedy and random walk, GDM-Net++ performs by 278% and 852% better in terms of earned rewards, while outperforming them by 34% and 155%, respectively, in terms of the precision of gas distribution modeling across unseen 2D test cases. Our approach can also gracefully handle 2D GDM scenarios where the distribution is consistently affected by wind.

In a gas distribution mapping (GDM) task, the objective of a mobile robot is to map the gas concentrations of an airborne chemical over a region of interest using onboard sensing. Given the limited battery budget available to the robot, covering the entire area to measure gas concentrations at every location might be infeasible. Assuming that the robot only has a budget for b meters of travel, in the rest of the locations, gas concentrations can be inferred using a supervised machine learning technique, namely the Gaussian Process (GP). In this paper, we propose a novel technique that combines deep reinforcement learning and GP regression to find an effective policy for GDM. We have implemented the proposed technique in Python within a 16×16 4-connected plane. We have used six types of Gaussian plumes to validate our presented approach. Compared to two popular baselines, our approach outperforms greedy and random exploration by 62% and 151% in terms of earned rewards, while outperforming them by 47% and 345%, respectively, in terms of the precision of gas distribution modeling in all test cases without obstacles. Our approach also improves the coverage of the exploration while consequently reducing the uncertainty in the prediction.

Disease control is paramount in modern agriculture to ensure optimal yield. Monitoring the spread of crop diseases is crucial for effective control measures. Traditional methods involve uniform pesticide spraying across entire fields, which can be inefficient and environmentally harmful. In this paper, we propose an intelligent solution employing mobile robots equipped with predictive AI techniques for disease monitoring and targeted intervention. These robots strategically visit select locations within the field, guided by a convolutional and recurrent neural network model trained on limited data to predict disease spread. We introduce a novel weighted path planning algorithm to optimize robot movement within the field considering disease risk and battery constraints. Our approach is implemented in the WaterBerry benchmark, an open-source platform for agricultural robotics. Experimental results demonstrate the efficacy of our technique, showcasing improved prediction accuracy and operational efficiency compared to baseline methods.

This paper introduces Confidence Guided Path-planning (CGP), an algorithm for planning the path of mobile sensor nodes with the goal to increase confidence in the accuracy of the estimated model at any time point in the data collection process. The approach employs a local estimator based on a Gaussian process regressor and takes advantage of the uncertainty estimation to guide the sensor to areas of lower confidence. In an experimental study comparing CGP with systematic lawnmower-type exploration and random waypoint movement, we found that CGP achieves better scores than both during most of the exploration process, being outperformed only by a fully completed systematic exploration. We also found that, as an emergent property of pursuing higher confidence, CGP achieves good coverage of the area of interest. The proposed algorithm has wide applications in precision agriculture, wildlife tracking, and road monitoring, where exhaustive coverage is not feasible.

Locating the source of harmful, flammable, or polluting gas leaks is an important task in many practical scenarios. A recently proposed localization approach is to use a mobile robot equipped with a chemical sensor. The localization algorithm guides the movement of the robot based on the previous observations, with the objective of reaching the source as quickly as possible. In this paper, we propose an approach where the robot policy is represented by a neural network combining convolutional and LSTM layers. The approach relies on a gas dispersion model that takes into account obstacles, wind direction, and molecular movement. We found that the trained model provides a 47.34% higher success rate in finding the gas source than an existing greedy approach on test cases with unseen gas plumes and random obstacles.

The movement of a mobile sensor has a critical impact on the information gathered from the area of interest, as well as the quality of the estimate that a model can build from the collected information at any moment in time. Both systematic exploration models, which make the sensor move in regular patterns, and random movement models have specific advantages. There is less research concerning models that are positioned between these two extremes. In this paper, we propose Grid Limited Randomness (GLR), a family of path planning algorithms based on sampling waypoints from a grid of a specific resolution. We propose three variations differentiated by the order in which the mobile sensor visits these waypoints: new samples added to the end of the path (GLR-EOP), smallest detour (GLR-SD), and the shortest path as approximated by Christofides' algorithm. An extensive simulation study in the Waterberry Farms benchmark shows that the GLR variations offer benefits that, in specific circumstances, make them preferable to both fully random and fully systematic exploration paths.

Large Language Models (LLMs) are progressively being utilized as machine learning services and interface tools for various applications. However, the security implications of LLMs, particularly in relation to adversarial and Trojan attacks, remain insufficiently examined. In this paper, we propose TrojLLM, an automatic and black-box framework to effectively generate universal and stealthy triggers. When these triggers are incorporated into the input data, the LLMs' outputs can be maliciously manipulated. Moreover, the framework also supports embedding Trojans within discrete prompts, enhancing the overall effectiveness and precision of the triggers' attacks. Specifically, we propose a trigger discovery algorithm for generating universal triggers for various inputs by querying victim LLM-based APIs using few-shot data samples. Furthermore, we introduce a novel progressive Trojan poisoning algorithm designed to generate poisoned prompts that retain efficacy and transferability across a diverse range of models. Our experiments and results demonstrate TrojLLM's capacity to effectively insert Trojans into text prompts in real-world black-box LLM APIs including GPT-3.5 and GPT-4, while maintaining exceptional performance on clean test sets. Our work sheds light on the potential security risks in current models and offers a potential defensive approach. The source code of TrojLLM is available at https://github.com/UCF-ML-Research/TrojLLM.

In today's era of automation, mobile robots are being used for collecting meaningful information about an ambient phenomenon such as temperature or moisture distribution in an agricultural field. Most of the studies in the literature assume that the underlying information field is Gaussian, and therefore, Gaussian Process (GP)-based models are extremely popular. Furthermore, we have found that due to the inherent computational complexity of such naive GP-based techniques, most studies in the literature do not scale well beyond small-size environments, i.e., where the number of informative points$n < 1000$. These render such a predictive model more or less useless in many practical applications. In this paper, we posit that a different technique, Generative Adversarial Network-based inpainting, for robotic information gathering can be useful. The state-of-art inpainting techniques 1) do not assume that the underlying data is Gaussian, and 2) easily scale to$n\gg 1000$. Thus, they eliminate the two bottlenecks posed by the GP-based solutions. We have tested our hypothesis on a synthetic and a real-world crop dataset. Results show that while the inpainting technique easily scales to$1024\times 1024$, GP-based predictions cannot. On the other hand, their solution qualities are shown to be comparable.

Multi-robot teams are becoming an increasingly popular approach for information gathering in large geographic areas, with applications in precision agriculture, surveying the aftermath of natural disasters or tracking pollution. These robot teams are often assembled from untrusted devices not owned by the user, making the maintenance of the integrity of the collected samples an important challenge. Furthermore, such robots often operate under conditions of opportunistic, or periodic connectivity and are limited in their energy budget and computational power. In this paper, we propose algorithms that build on blockchain technology to address the data integrity problem, but also take into account the limitations of the robots' resources and communication. We evaluate the proposed algorithms along the perspective of the tradeoffs between data integrity, model accuracy, and time consumption.

Several recent papers introduced techniques to adjust the attributes of human faces generated by unconditional GANs such as StyleGAN. Despite efforts to disentangle the attributes, a request to change one attribute often triggers unwanted changes to other attributes as well. More importantly, in some cases, a human observer would not recognize the edited face to belong to the same person. We propose an approach where a neural network takes as input the latent encoding of a face and the desired attribute changes and outputs the latent space encoding of the edited image. The network is trained offline using unsupervised data, with training labels generated by an off-the-shelf attribute classifier. The desired attribute changes and conservation laws, such as identity maintenance, are encoded in the training loss. The number of attributes the mapper can simultaneously modify is only limited by the attributes available to the classifier – we trained a network that handles 35 attributes, more than any previous approach. As no optimization is performed at deployment time, the computation time is negligible, allowing real-time attribute editing. Qualitative and quantitative comparisons with the current state-of-the-art show our method is better at conserving the identity of the face and restricting changes to the requested attributes.

Blockchain-powered smart systems deployed in different industrial applications promise operational efficiencies and improved yields, while significantly mitigating cybersecurity risks. Tradeoffs between availability and security arise at implementation, however, triggered by the additional resources (e.g., memory and computation) required by blockchain-enabled hosts. This paper applies an energy-reducing algorithmic engineering technique for Merkle Tree (MT) root calculations and the Proof of Work (PoW) algorithm, two principal elements of blockchain computations, as a means to preserve the promised security benefits but with less compromise to system availability. Using pyRAPL, a python library to measure the energy consumption of a computation, we experiment with both the standard and energy-reduced implementations of both algorithms for different input sizes. Our results show that up to 98% reduction in energy consumption is possible within the blockchain’s MT construction module, with the benefits typically increasing with larger input sizes. For the PoW algorithm, our results show up to 20% reduction in energy consumption, with the benefits being lower for higher difficulty levels. The proposed energy-reducing technique is also applicable to other key elements of blockchain computations, potentially affording even “greener” blockchain-powered systems than implied by only the results obtained thus far on the MT and PoW algorithms.

The daily activities performed by a disabled or elderly person can be monitored by a smart environment, and the acquired data can be used to learn a predictive model of user behavior. To speed up the learning, several researchers designed collaborative learning systems that use data from multiple users. However, disclosing the daily activities of an elderly or disabled user raises privacy concerns.In this paper, we use state-of-the-art deep neural network-based techniques to learn predictive human activity models in the local, centralized, and federated learning settings. A novel aspect of our work is that we carefully track the temporal evolution of the data available to the learner and the data shared by the user. In contrast to previous work where users shared all their data with the centralized learner, we consider users that aim to preserve their privacy. Thus, they choose between approaches in order to achieve their goals of predictive accuracy while minimizing the shared data. To help users make decisions before disclosing any data, we use machine learning to predict the degree to which a user would benefit from collaborative learning. We validate our approaches on real-world data.

We study the problem of information sampling of an ambient phenomenon using a group of mobile robots. Autonomous robots are being deployed for various applications such as precision agriculture, search-and-rescue, among others. These robots are usually equipped with sensors and tasked with collecting maximal information for further data processing and decision making. The studied problem is proved to be NP-Hard in the literature. To solve the stated problem approximately, we employ a multi-agent deep reinforcement learning framework and use the concepts of mean field games to potentially scale the solution to larger multi-robot systems. Simulation results show that our presented technique easily scales to 10 robots in a 19 × 19 grid environment, while consistently sampling useful information.

Blockchain-powered smart systems deployed in different industrial applications promise operational efficiencies and improved yields, while mitigating significant cybersecurity risks pertaining to the main application. Associated tradeoffs between availability and security arise at implementation, however, triggered by the additional resources (e.g., memory, computation) required by each blockchain-enabled host. This paper applies an energy-reducing algorithmic engineering technique for Merkle Tree root calculations, a principal element of blockchain computations, as a means to preserve the promised security benefits but with less compromise to system availability. Using pyRAPL, a python library to measure computational energy, we experiment with both the standard and energy-reduced implementations of the Merkle Tree for different input sizes (in bytes). Our results show up to 98% reduction in energy consumption is possible within the blockchain's Merkle Tree construction module, such reductions typically increasing with larger input sizes. The proposed energy-reducing technique is similarly applicable to other key elements of blockchain computations, potentially affording even “greener” blockchain-powered systems than implied by only the Merkle Tree results obtained thus far.

We propose a novel method for automatic object recoloring based on Generative Adversarial Networks (GANs). The user can simply give commands of the form recolortowhich will be executed without any need of manual edit. Our approach takes advantage of pre-trained object detectors and saliency mask segmentation networks. The segmented mask of the given object along with the target color and the original image form the input to the GAN. The use of cycle consistency loss ensures the realistic look of the results. To our best knowledge, this is the first algorithm where the automatic recoloring is only limited by the ability of the mask extractor to map a natural language tag to a specific object in the image (several hundred object types at the time of this writing). For a performance comparison, we also adapted other state of the art methods to perform this task. We found that our method had consistently yielded qualitatively better recoloring results.

Many elderly individuals have physical restrictions that require the use of a walker to maintain stability while walking. In addition, many of these individuals also have age-related visual impairments that make it difficult to avoid obstacles in unfamiliar environments. To help such users navigate their environment faster, safer and more easily, we propose a smart walker augmented with a collection of ultrasonic sensors as well as a camera. The data collected by the sensors is processed using echo-location based obstacle detection algorithms and deep neural networks based object detection algorithms, respectively. The system alerts the user to obstacles and guides her on a safe path through audio and haptic signals.

Recent research demonstrated that it is feasible to end-to-end train multi-task deep visuomotor policies for robotic manipulation using variations of learning from demonstration (LfD) and reinforcement learning (RL). In this paper, we extend the capabilities of end-to-end LfD architectures to object manipulation in clutter. We start by introducing a data augmentation procedure called Accept Synthetic Objects as Real (ASOR). Using ASOR we develop two network architectures: implicit attention ASOR-IA and explicit attention ASOR-EA. Both architectures use the same training data (demonstrations in uncluttered environments) as previous approaches. Experimental results show that ASOR-IA and ASOR-EA succeed in a significant fraction of trials in cluttered environments where previous approaches never succeed. In addition, we find that both ASOR-IA and ASOR-EA outperform previous approaches even in uncluttered environments, with ASOR-EA performing better even in clutter compared to the previous best baseline in an uncluttered environment.

Many cooperative multi-agent problems require agents to learn individual tasks while contributing to the collective success of the group. This is a challenging task for current state-of-the-art multi-agent reinforcement algorithms that are designed to maximize either the global reward of the team or the individual local rewards. The problem is exacerbated when either of the rewards is sparse leading to unstable learning. To address this problem, we present Decomposed Multi-Agent Deep Deterministic Policy Gradient (DE-MADDPG): a novel cooperative multi-agent reinforcement learning framework that simultaneously learns to maximize the global and local rewards. We evaluate our solution on the defensive escort team problem and show that our solution achieves better and more stable performance than the direct adaptation of the MADDPG algorithm.

The smart home concept can significantly benefit from predictive models that take proactive management operations on home actuators, based on users' behavior evaluation. In this paper, we use a small-scale physical model, the ScaledHome-2 testbed, to experiment with the evolution of measurements in a suburban home under different environmental scenarios. We start from the observation that, for a home to become smart, in addition to IoT sensors and actuators, we also need a predictive model of how actions taken by inhabitants and home actuators affect the internal environment of the home, reflected in the sensor readings. In this paper, we propose a technique to create such a predictive model through machine learning in various simulated weather scenarios. This paper also contributes to the literature in the field by quantitatively comparing several machine learning algorithms (K-nearest neighbor, regression trees, Support Vector Machine regression, and Long Short Term Memory deep neural networks) in their ability to create accurate and generalizable predictive models for smart homes.

A rich amount of literature is available for solving the problem of finding equilibrium strategies in two-player security games that harness the power of integer linear programming (ILP). However, in practice, most security games are accurately modeled with multiple agents where ILP methods either fail to find the optimal solution or the state space is large enough making ILP methods an impractical solution. In this paper, we consider a multi-agent security game setting and propose MultiOptGrad: a novel deep reinforcement learning-based solution to learn distributed optimal policies for defenders. Additionally, using MultiOptGrad we built an reinforcement learning framework for robotic bodyguards that recommend deployment strategies for them in a coordinate system. To demonstrate the effectiveness of our proposed solution, we consider an urban security game where a team of robotic bodyguards are protecting a VIP from physical assault in the presence of neutral and/or adversarial bystanders. Our empirical analysis has shown that MultiOptGrad outperformed quadrant load-balancing (QLB): a hand-engineered technique for solving the VIP protection problem.

Several recent studies have demonstrated the promise of deep visuomotor policies for robot manipulator control. Despite impressive progress, these systems are known to be vulnerable to physical disturbances, such as accidental or adversarial bumps that make them drop the manipulated object. They also tend to be distracted by visual disturbances such as objects moving in the robot's field of view, even if the disturbance does not physically prevent the execution of the task. In this paper, we propose an approach for augmenting a deep visuomotor policy trained through demonstrations with Task Focused visual Attention (TFA). The manipulation task is specified with a natural language text such as "move the red bowl to the left". This allows the visual attention component to concentrate on the current object that the robot needs to manipulate. We show that even in benign environments, the TFA allows the policy to consistently outperform a variant with no attention mechanism. More importantly, the new policy is significantly more robust: it regularly recovers from severe physical disturbances (such as bumps causing it to drop the object) from which the baseline policy, i.e. with no visual attention, almost never recovers. In addition, we show that the proposed policy performs correctly in the presence of a wide class of visual disturbances, exhibiting a behavior reminiscent of human selective visual attention experiments.

Recent advances in IoT sensors and actuators and smart home controllers allow us to collect real-time information about the state of the home and take intelligent actions that maximize the user's goals with respect to comfort, convenience, environmental awareness and cost. While thermal comfort is one of the primary concerns of many users, many homes use a very simple, energy inefficient approach that blankets the home with constant temperature air conditioning. Such systems do not take advantage of more energy efficient and environment friendly natural ways to manage the temperature, such as opening and closing windows, window shades and interior doors. In this paper we develop a deep neural network based model that predicts the temperature in various rooms of the home function of the state of the actuators. We also describe a scaled model of a four room home which allows us to control the doors and windows and collect data using IoT devices. We train and validate our temperature models on both data collected from the scaled model as well as from publicly available datasets from two real-world smart homes.

Visually impaired individuals often employ canes or guide dogs to help navigate complex environments. Individuals who are both visually and mobility impaired encounter greater difficulty, since conventional aids do not integrate well with walkers or rollators. In this paper we propose a smart walker architecture that is augmented with depth-sensing cameras that detect and recognize obstacles that may endanger the user, as well as obtain their distance from the user. This information is conveyed to the user through a haptic interface on the walker. Our design helps protect the user from colliding with obstacles in his/her path. In comparison with traditional mobility methods, the smart walker allows the user to navigate the environment faster and with less physical and cognitive load. Compared to previous designs, our approach completes this task at a significantly lower cost per unit.

Four legged walkers are used by many elderly persons to retain mobility. They are also used by patients recovering from leg injuries to facilitate rehabilitation. Unfortunately, these walkers are also associated with many injuries, some of which are caused by incorrect use. In this paper, we describe a walker augmented with IoT sensors which continuously monitors the weight distribution on the legs of the walker. We describe an approach where this data stream is processed by a deep neural network based classifier, which learns to recognize dangerous use patterns that can lead to falls and injury. The classifier is trained by providing examples of unsafe use, thus eliminating the costly engineering necessary to customize the algorithm to the specific user and walker. By alerting the user in real time about unsafe use patterns, the user can learn the correct and safe use of the walker.

Few-shot or one-shot learning of classifiers requires a significant inductive bias towards the type of task to be learned. One way to acquire this is by meta-learning on tasks similar to the target task. In this paper, we propose UMTRA, an algorithm that performs unsupervised, model-agnostic meta-learning for classification tasks. The meta-learning step of UMTRA is performed on a flat collection of unlabeled images. While we assume that these images can be grouped into a diverse set of classes and are relevant to the target task, no explicit information about the classes or any labels are needed. UMTRA uses random sampling and augmentation to create synthetic training tasks for meta-learning phase. Labels are only needed at the final target task learning step, and they can be as little as one sample per class. On the Omniglot and Mini-Imagenet few-shot learning benchmarks, UMTRA outperforms every tested approach based on unsupervised learning of representations, while alternating for the best performance with the recent CACTUs algorithm. Compared to supervised model-agnostic meta-learning approaches, UMTRA trades off some classification accuracy for a reduction in the required labels of several orders of magnitude.

Robots assisting the disabled or elderly must perform complex manipulation tasks and must adapt to the home environment and preferences of their user. Learning from demonstration is a promising choice, that would allow the non-technical user to teach the robot different tasks. However, collecting demonstrations in the home environment of a disabled user is time consuming, disruptive to the comfort of the user, and presents safety challenges. It would be desirable to perform the demonstrations in a virtual environment. In this paper we describe a solution to the challenging problem of behavior transfer from virtual demonstration to a physical robot. The virtual demonstrations are used to train a deep neural network based controller, which is using a Long Short Term Memory (LSTM) recurrent neural network to generate trajectories. The training process uses a Mixture Density Network (MDN) to calculate an error signal suitable for the multimodal nature of demonstrations. The controller learned in the virtual environment is transferred to a physical robot (a Rethink Robotics Baxter). An off-the-shelf vision component is used to substitute for geometric knowledge available in the simulation and an inverse kinematics module is used to allow the Baxter to enact the trajectory. Our experimental studies validate the three contributions of the paper: (1) the controller learned from virtual demonstrations can be used to successfully perform the manipulation tasks on a physical robot, (2) the LSTM+MDN architectural choice outperforms other choices, such as the use of feedforward networks and mean-squared error based training signals and (3) allowing imperfect demonstrations in the training set also allows the controller to learn how to correct its manipulation mistakes.

This paper focuses on the application of wireless sensor networks (WSNs) with unmanned aerial vehicle (UAV) for animal tracking problem. The goal of this application is to monitor the target animals in large wild areas without any attachment devices. The WSN includes clusters of sensor nodes and a single UAV that acts as a mobile sink and visits the clusters. We propose a model predictive control (MPC) method that is used to guide the UAV in planning its path. We first build a prediction model to learn the animal appearance patterns from the sensed historical data. Then, based on the real-time predicted animal distributions, we introduce a path planning approach for the UAV that reduces message delay by maximizing the collected rewards. The experimental results show that our approach outperforms the greedy and traveling salesmen problem-based path planning heuristics in terms of collected value of information. We also discuss the results of other performance metrics involving message delay and percentage of events collected.

We propose a technique for multi-task learning from demonstration that trains the controller of a low-cost robotic arm to accomplish several complex picking and placing tasks, as well as non-prehensile manipulation. The controller is a recurrent neural network using raw images as input and generating robot arm trajectories, with the parameters shared across the tasks. The controller also combines VAE-GAN-based reconstruction with autoregressive multimodal action prediction. Our results demonstrate that it is possible to learn complex manipulation tasks, such as picking up a towel, wiping an object, and depositing the towel to its previous position, entirely from raw images with direct behavior cloning. We show that weight sharing and reconstruction-based regularization substantially improve generalization and robustness, and training on multiple tasks simultaneously increases the success rate on all tasks.

In this paper, we focus on a taxi dispatch system with the help of auxiliary models that predict future demand and destination. We build two different neural networks for learning taxi demand and destination distribution patterns based on historical data. The trained models can predict taxi demand and destination for any area in a city at a future time. Our proposed dispatch system relies on the predictions of the previous models and is designed not only to minimize the waiting time of passengers, but also to assign the taxis to passengers in a way to minimize the idle driving distances of taxis. In order to achieve this, we balance future taxi supply-demand over the city by solving a mixed-integer program (MIP). We validate our dispatch system as well as the prediction models using a dataset of taxi trips in the New York City.

Predicting taxi demand throughout a city can help to organize the taxi fleet and minimize the wait-time for passengers and drivers. In this paper, we propose a sequence learning model that can predict future taxi requests in each area of a city based on the recent demand and other relevant information. Remembering information from the past is critical here, since taxi requests in the future are correlated with information about actions that happened in the past. For example, someone who requests a taxi to a shopping center, may also request a taxi to return home after few hours. We use one of the best sequence learning methods, long short term memory that has a gating mechanism to store the relevant information for future use. We evaluate our method on a data set of taxi requests in New York City by dividing the city into small areas and predicting the demand in each area. We show that this approach outperforms other prediction methods, such as feed-forward neural networks. In addition, we show how adding other relevant information, such as weather, time, and drop-offs affects the results.

We consider underwater multi-modal wireless sensor networks (UWSNs) suitable for applications on submarine surveillance and monitoring, where nodes offload data to a mobile autonomous underwater vehicle (AUV) via optical technology, and coordinate using acoustic communication. Sensed data are associated with a value, decaying in time. In this scenario, we address the problem of finding the path of the AUV so that the Value of Information (VoI) of the data delivered to a sink on the surface is maximized. We define a Greedy and Adaptive AUV Path-finding (GAAP) heuristic that drives the AUV to collect data from nodes depending on the VoI of their data. For benchmarking the performance of AUV path-finding heuristics, we define an integer linear programming (ILP) formulation that accurately models the considered scenario, deriving a path that drives the AUV to collect and deliver data with the maximum VoI. In our experiments GAAP consistently delivers more than 80 percent of the theoretical maximum VoI determined by the ILP model. We also compare the performance of GAAP with that of other strategies for driving the AUV among sensing nodes, namely, random paths, TSP-based paths and a “lawn mower”-like strategy. Our results show that GAAP always outperforms every other heuristic in terms of delivered VoI, also obtaining higher energy efficiency.

Many companies within the Internet of Things (IoT) sector rely on the personal data of users to deliver and monetize their services, creating a high demand for personal information. A user can be seen as making a series of transactions, each involving the exchange of personal data for a service. In this paper, we argue that privacy can be described quantitatively, using the game- theoretic concept of value of information (VoI), enabling us to assess whether each exchange is an advantageous one for the user. We introduce PrivacyGate, an extension to the Android operating system built for the purpose of studying privacy of IoT transactions. An example study, and its initial results, are provided to illustrate its capabilities.

In this paper, we focus on an application of recurrent neural networks for learning a model that predicts taxi demand based on the requests in the past. A model that can learn time series data is necessary here since taxi requests in the future relate to the requests in the past. For instance, someone who requests a taxi to a movie theater, may also request a taxi to return home after few hours. We use Long Short Term Memory (LSTM), one of the best models for learning time series data. For training the network, we encode the historical taxi requests from the official New York City taxi trip dataset and add date, day of the week and time as impacting factors. Experimental results show that our approach outperforms the prediction heuristics based on feed-forward neural networks and naive statistic average.

In Underwater Sensor Networks (UWSNs) with high volume of data recording activity, a mobile sink such as a Autonomous Underwater Vehicle (AUV) can be used to offload data from the sensor nodes. When the AUV approaches the underwater node, it can use high data rate optical communication. However, the data is not considered delivered when it was transferred from the sensor node to the AUV, but when the AUV had resurfaced and transferred the data to the sink. If the data is not time sensitive, it is sufficient for the AUV to resurface only once at the end of its data collection path. However, for time-sensitive data, it is more advantageous for the AUV to resurface multiple times during its path, and upload the data collected since the previous resurfacing. Thus, a resurfacing schedule needs to complement the path planning process. In this paper we are using the metric of Value of Information (VoI) as the optimization criteria to capture the time- sensitive nature of collected information. We propose a genetic algorithm based approach to determine the resurfacing schedule for an AUV which is already provided with the sequence of nodes to be visited.

Picking up an object with a wheelchair mounted robotic arm can be decomposed into a wheelchair navigation task designed to position the robotic arm such that the object is “easy to reach”, and the actual grasp performed by the robotic arm. A convenient definition of the notion of ease of reach can be given by creating a score (ERS) that relies on the number of distinct ways the object can be picked up from a given location. Unfortunately, the accurate calculation of ERS must rely on repeating the path planning process for every candidate position and grasp type, in the presence of obstacles. In this paper we use the bootstrap aggregation over hand-crafted, domain specific features to learn a model for the estimation of ERS. In a simulation study, we show that the estimated ERS closely matches the actual value and the speed of estimation is fast enough for real-time operation, even in the presence of a large number of obstacles in the scene.

In a wireless sensor network, virtual coordinates provide most of the advantages of geographic routing strategies without actually relying on the location information of the nodes. Using a mobile sink provides advantages such as distributing energy consumption throughout the network. However, nodes need to be updated about the new virtual coordinate of the mobile sink as it moves. In this paper, we propose Circular Update-Directional Virtual Coordinate Routing (CU-DVCR), an algorithm specialized in routing towards a mobile sink in virtual coordinates. Through a set of experimental studies we show that CU-DVCR consumes less energy compared to alternative algorithms while providing comparable performance.

Underwater Sensor Networks (UWSNs) provide valuable data for research studies and underwater monitoring and protection. UWSNs need to overcome the handicap that high data rate wireless transmissions are not available underwater. Acoustic communications are used as a medium but they are only good for transmitting e.g. signalling information. Autonomous Underwater Vehicles (AUVs) can serve as mobile sinks that gather and deliver larger amounts of data from the underwater sensor network nodes. Value of Information (VoI) is a data tag that encodes the importance and time-based-relevance of a data chunk residing at a sensor node. VoI, therefore, can serve as a heuristic for path planning and prioritizing data retrieval from nodes. The novelty of this paper lies in providing algorithms which schedule multiple mobile sinks (AUVs) for data retrieval from nodes while maximizing the retrieved VoI. The class of algorithms discussed are based on greedy heuristics.

In this paper, we focus on an application of wireless sensor networks (WSNs) with unmanned aerial vehicle (UAV). The aim of the application is to detect the locations of endangered species in large-scale wildlife areas or monitor movement of animals without any attachment devices. We first define the mathematical model of the animal monitoring problem in terms of the value of information (VoI) and rewards. We design a network model including clusters of sensor nodes and a single UAV that acts as a mobile sink and visits the clusters. We propose a path planning approach based on a Markov decision process (MDP) model that maximizes the VoI while reducing message delays. We used real-world movement dataset of zebras. Simulation results show that our approach outperforms greedy and random heuristics as well as the path planning based on the solution of the traveling salesman problem.

Geographical routing can provide significant advantages in wireless sensor networks. However in many sensor networks, it is difficult or costly to find the exact location of the nodes. The virtual coordinate techniques allow a network to acquire a coordinate system without relying on geographical location. In this paper, we describe MS-DVCR, an extension of a state-of-the-art virtual coordinate routing protocol (DVCR) with the ability to route towards a mobile sink. We describe the design principles and implementation of the proposed protocol and through an experimental study, we show that it matches the performance of a simple extension of DVCR for mobile sinks while providing a significantly lower energy consumption.

This paper considers underwater wireless sensor networks (UWSNs) for submarine surveillance and monitoring. Nodes produce data with an associated value, decaying in time. An autonomous underwater vehicle (AUV) is sent to retrieve information from the nodes, through optical communication, and periodically emerges to deliver the collected data to a sink, located on the surface or onshore. Our objective is to determine a collection path for the AUV so that the Value of Information (VoI) of the data delivered to the sink is maximized. To this purpose, we first define an Integer Linear Programming (ILP) model for path planning that considers realistic data communication rates, distances, and surfacing constraints. We then define the first heuristic for path finding that is adaptive to the occurrence of new events, relying only on acoustic communication for exchanging short control messages. Our Greedy and Adaptive AUV Path-finding (GAAP) heuristic drives the AUV to collect packets from nodes to maximize the VoI of the delivered data. We compare the VoI of data obtained by running the optimum solution derived by the ILP model to that obtained from running GAAP over UWSNs with realistic and desirable size. In our experiments GAAP consistently delivers more than 80% of the theoretical maximum VoI determined by the ILP model.

We consider an underwater wireless sensor network where baseline communication happens over acoustic, multi-hop routes from the underwater nodes to an on-shore station. The data collected by the nodes greatly exceeds the baseline communication capability. At best, the nodes can transmit digests of their full observations. In order for the sink to receive all sensed data, an autonomous underwater vehicle (AUV) is sent to each node for collecting data over short-distance, high data rate optical connections. The AUV then offloads all collected information to the terrestrial station via wireless communication when it surfaces. The observations made by the nodes vary in size and urgency. The information they provide has an associated value. Given a path of the AUV, we design scheduling strategies for the nodes to decide when and how much information (i.e., which digest) to transmit via acoustic routes so that the value of information reaching the terrestrial station is maximized. These strategies are compared via simulations on realistic scenarios. Our results show that scheduling algorithms that are able to locally estimate the value of information of a data digest provide the delivery of data with a significantly higher value of information. In contrast, uninformed algorithms, i.e., strategies that do not consider the value of information at the node level, provide only a marginal increase over the benchmark case of using only the AUV for data collection.

We have developed a multi-agent negotiation system to distribute decision making in cognitive radio networks through argumentation. The challenge in wireless network negotiation is to efficiently exchange information to facilitate a deal without incurring excessive communication overhead or indeterminate negotiation time. Our goal is to improve both total network throughput and the number of total supported connections. We detail a set of rules, a protocol, and a compact set of messages to conduct these negotiations and complete in finite time and with little overhead. We describe our simulation environment and present results of an illustrative scenario with various conditions. This scenario includes the ability of an agent to assert high priority, possibly triggering a downgrade of an existing, non-priority connection to a slower rate in order to accommodate more connections. We compare our system's total network throughput, number of connections, and request satisfaction score to several baselines with various levels of reconsideration and conclude that our system outperforms these other approaches in all metrics.

This paper proposes a reporting decision protocol called IVE (for Information Value - Energy tradeoff), where individual nodes of an intruder tracking sensor network make decisions about the transmission of information chunks. Instead of trying to achieve raw data metrics (such as total transmitted data) the protocol aims to optimize the value of information (VoI) maintained by the customer. To achieve this, the nodes will need to perform inferences about the behavior of other nodes and the customer, such that the nodes do not need to send information which the customer already received from other sources or information which it can guess based on previous data. A simulation study compares the performance of the IVE protocol with the current state of the art of on-demand periodic reporting.

Sensor networks are distributed systems where nodes embedded in the environment collect readings through their sensors and transmit data to customers. The overall goal of these systems can be stated as maximizing a metric of the sensing quality while limiting the consumption of a set of scarce resources. In this paper we consider an intruder detection and tracking system where the sensing quality is a metric of the pragmatic value of the information provided by the network. This metric depends not only on the quantity and accuracy of information, but also on when and how the customers will use this information. We design a system which adapts its information transmission to the disruptive decisions made by the user, including a consideration for the cost of incorrect decisions.

The worst case scenario for the life cycle of a sensor network is the fragmentation of a network which still has many functional and well-powered nodes. The loss of connectivity renders even the functional nodes useless as the nodes are not able to transmit their observations to the sink. A well-engineered sensor network will not fragment due to the energy consumption occurring during normal functioning. However, catastrophic events, which unpredictably destroy a large subset of the nodes, can transform a well engineered network into a heavily unbalanced one. Very often, even if the network is not yet fragmented, the connectivity is relying on one or more bridge nodes, which survived the catastrophic event accidentally. If the network operates as before, the bridge nodes will soon exhaust their power resources by having to route an unexpectedly large number of packets. This paper describes the Bridge Protection Algorithm (BPA), a combination of techniques which, in response to a catastrophic event, change the behavior of a set of topologically important nodes in the network. These techniques protect the bridge node by letting some nodes take over some of the responsibilities of the sink. At the same time, they relieve some other overwhelmed nodes and prevent the apparition of additional bridge nodes. To achieve this, BPA sacrifices the length of some routes in order to distribute routes away from critical areas. Through a simulation study we show that the application of these techniques can significantly decrease the load of the nodes in the critical areas, while only minimally affecting the performance of the network.

A large part of the energy budget of traditional sensor networks is consumed by the hop-by-hop routing of the collected information to the static sink. In many applications it is possible to replace the static sink with one or more mobile sinks that move in a sensor field and collect the data through one-hop transmissions. This greatly reduces the power consumption of the nodes, which can be further reduced by choosing the appropriate moment of transmission. In general, the transmission energy increases quickly with the distance, and thus it makes sense for the nodes to transmit when one of the mobile sinks is in close proximity. Seeing the node as an autonomous agent, it needs to choose its actions of transmitting or buffering the collected data based on what it knows about the environment and its predictions about the future. The sensor agent needs to appropriately balance the following two objectives: the maximization of the utility of the collected and transmitted data and the minimization of the energy expenditure. We introduce the cummulative policy penalty as an expression of this interdependent pair of requirements. As a baseline, we describe a graph-theory-based approach for calculating the optimal policy in a complete knowledge setting. Then, we describe and compare three heuristics based on different principles (imitation of human decision making, stochastic transmission and constant risk). We compare the proposed approaches in an experimental study under a variety of scenarios.

Many sensor networks are deployed to detect and track intruders. If the existence and location of sensor nodes is disclosed to the opponent, the nodes can be easily disabled or compromised. Wireless transmissions in the presence of the opponent are an important source of disclosure. In this paper, we first describe a way to quantify the stealthiness of the sensor node, with a numerical stealthiness metric. Then, we introduce a local model based dissemination protocol, try and bounce (TAB) which takes into account stealth considerations while reporting and forwarding observation reports.

Autonomous rechargeable sensor networks are becoming a feasible solution to many real world applications. In this paper, we propose a Uniform Sensing Protocol for autonomous rechargeable sensor networks. Our protocol aims to provide uniformly distributed sensing throughout the entire life-time of the network, thus increasing the overall network reliability. It considers the amount of available energy in the environment as well as the probability of encountering a specific number of threats. Using these parameters, each node estimates its own active period, such that uniform sensing is established. We compare the performance of our protocol with static and dynamic active time slot approaches. The simulation results show that the Uniform Sensing Protocol generates fewer failures and has a significantly longer mean time to failure than the other two schemes.

The urban environment of the early 21st century contains a large number of consumer devices under private and organizational ownership. Many of these devices contain sensors as well as communication devices. However, most of these sensors are only used for purposes internal to the device. By interconnecting these sensors we can obtain a network which can serve important societal goals. The technological challenges of interconnecting these sensors are relatively minor. The main problem is the human aspect: why would the owners of the sensors offer their readings for public use? In fact, privacy considerations might advise the exact opposite. Such networks will not be accepted unless every owner is motivated to allow the participation of its devices. In this paper we describe an architecture which enables such a system by the formal model of motivational gradients. The original source of motivational gradients are always humans or organizations; however, nodes acting as autonomous agents can negotiate motivational microgradients based on the original macrogradient introduced by humans. We evaluate the networking and computer-human interaction aspects of the proposed architecture.

Abstract We report on a study in which 12 different paradigms were used to implement agents acting in an environment which borrows elements from artificial life and multi‐player strategy games. In choosing the paradigms we strived to maintain a balance between high‐level, logic‐based approaches and low‐level, physics‐oriented models; between imperative programming, declarative approaches and ‘learning from basics’; between anthropomorphic or biologically inspired models on one hand and pragmatic, performance‐oriented approaches on the other. We have found that the choice of the paradigm determines the software development process and requires a different set of skills from the developers. In terms of raw performance, we found that the best performing paradigms were those which (a) allowed the knowledge of human experts to be explicitly transferred to the agent and (b) allowed the integration of well‐known, high‐performance algorithms. We have found that maintaining a commitment to the chosen paradigm can be difficult; there is a strong temptation to offer shallow fixes to perceived performance problems through a ‘flight into heuristics’. Our experience is that a development process without the discipline enforced by a central paradigm leads to agents which are a random collection of heuristics whose interactions are not clearly understood. Although far from providing a definitive verdict on the benefits of the different paradigms, our study provided a good insight into what kind of conceptual, technical or organizational problems would a development team face depending on their choice of agent paradigm. Copyright © 2007 John Wiley & Sons, Ltd.

Abstract Mobile sinks can significantly extend the lifetime of a sensor network by eliminating the need for expensive hop‐by‐hop routing. However, a sensor node might not always have a mobile sink in transmission range, or the mobile sink might be so far that the data transmission would be very expensive. In the latter case, the sensor node needs to make a decision whether it should send the data now, or take the risk to wait for a more favorable occasion. Making the right decisions in thistransmission scheduling problemhas significant impact on the performance and lifetime of the node. In this paper, we investigate the fundamentals of the transmission scheduling problem for sensor networks with mobile sinks. We first develop a dynamic programming‐based optimal algorithm for the case when the mobility of the sinks is known in advance. Then, we describe two decision theoretic algorithms which use only probabilistic models learned from the history of interaction with the mobile sinks, and do not require knowledge about their future mobility patterns. The first algorithm uses Markov Decision Processes with states without history information, while the second algorithm encodes some elements of the history into the state. Through a series of experiments, we show that the decision theoretic approaches significantly outperform naive heuristics, and can have a performance close to that of the optimal approach, without requiring an advance knowledge of the mobility. Copyright © 2007 John Wiley & Sons, Ltd.

In this paper, we report on a series of experiments involving the speedups obtainable with time-parallel simulation of wireless ad hoc networks. A mobile ad hoc network scenario involving the AODV and DSDV routing protocols was simulated. The results and the performance of the serial NS-2 simulator were compared to the time-parallel simulation method for wireless ad hoc networks, previously introduced by the authors. The results of the time-parallel simulation are approximations, and we find that there is a trade off between the precision of the simulation and the achievable speedup. However, it is possible to find compromises where a precision of the range of 95-98%, sufficient for most applications, can be obtained up to 10 times faster than the time needed by a serial simulation

In this paper we discuss the role of a broker in a market-oriented resource allocation model for large-scale heterogeneous systems. The simplified model is based upon a three party system, provider-broker-consumer. The allocation of resources is determined by their price, their utility to the consumer, and by the satisfaction of the consumer. The role of the broker is to add societal objectives to resource allocation algorithms and to mediate between greedy consumers and selfish providers. A simulation experiment was conducted to study the transient and the steady-state behavior of several performance measures, including the average consumer satisfaction, the average utility, and the hourly revenue.

Asymmetric links are common in wireless networks for a variety of physical, logical, operational, and legal considerations. An asymmetric link supports uni-directional communication between a pair of mobile stations and requires a set of relay stations for the transmission of packets in the other direction. We introduce a MAC layer protocol for wireless networks with Asymmetric links (AMAC). The MAC layer protocol requires fewer nodes to maintain silence during a transmission exchange than the protocols proposed in [1, 2]. We present a set of concepts and metrics characterizing the ability of a medium access control protocol to silence nodes which could cause collisions.

We introduce a scalable algorithm for time-parallel simulations of wireless ad hoc networks and report on our results. Our approach decomposes the simulation into overlapping temporal components; the individual components are computed using an unmodified sequential network simulator such as NS-2. Our algorithm is iterative and the accuracy of the results increases with the number of iterations. We find that the approach allows the simulation of ad hoc networks with a number of nodes larger than those feasible with sequential network simulators on single CPUs. The algorithm is scalable, we can simulate larger time intervals by increasing the number of virtual processors carrying out the simulation. We identify the parameters that can be investigated with the algorithm and report on the accuracy of our results and on the achieved simulation speedup

The global Internet is rich in commodity resources but scarce in specialized resources. We argue that a grid framework can achieve better performance if it separates management of commodity tasks from the management of the tasks requiring specialized resources. Assuming a relative homogeneity of the commodity resource providers, the determining factor of grid performance becomes the latency of entering into execution. This effectively transforms the resource allocation problem into a routing problem. We present an approach in which commodity tasks are distributed to the commodity service providers by request forwarding on the n-cycle overlay network. We provide algorithms for task allocation and for the maintenance of the overlay network. By ensuring that the algorithms use only narrow local information, the approach is easily scalable to millions of nodes. For task allocation algorithms in a commercial setting, fairness is of paramount importance. We investigate the properties of the proposed algorithms from the fairness point of view and show how adding several hops of random pre-walk to the algorithm can improve its fairness. Extensive simulations prove that the approach provides efficient task allocation on networks loaded up to 95% of their capacity.

Developing network protocols for mobile wireless systems is a complex task, and most of the existing simulator frameworks are not well suited for experimental development. The YAES simulation framework was specifically developed such that it allows the fast prototyping of networking protocols, and support real-time experimentation and refactoring. By providing a large set of abstractions and generic implementations, a number of frequently used techniques such as genetic algorithms or neural networks can be created in matter of minutes. Our experience shows that by requiring only Java programming skills which computer science and engineering students commonly possess, YAES can be a useful tool for classroom use, as well.This paper presents the considerations behind the YAES architecture and provides a description of the system. As a case study, we present the steps necessary for running experiments on the energy efficiency behavior of the Weighted Clustering Algorithm (WCA).

Convoy driving on public highways is a useful phenomena which increases the safety and the throughput of the highway. We present an approach through which a wireless Convoy Driving Device assists the driver in the task of deciding to join or leave a convoy, influencing the speed and formation of the convoy. Our approach handles complex situations like the merging and splitting of convoys, and it offers valuable lessons with applications for other cases of teamwork of mobile entities.

A computational grid is a complex system. The state space of a complex system is very large and it is infeasible to create a rigid infrastructure implementing optimal policies and strategies which take into account the current state of the system. An alternative to a rigid infrastructure is to base the system's reactions on logical inference, planning, and learning, the quintessential elements of an intelligent system. An intelligent grid is one where societal services exhibit intelligent behavior. A coordination service acting as a proxy on behalf of end users reacts to unforeseen events, plans how to carry out complex tasks, and learns from the history of the system. Various policies implemented by the societal services of an intelligent grid, such as brokerage and matchmaking, are based upon rules and facts gathered with the aid of a monitoring service. The question we address is how to construct intelligent computational grids which are truly scalable and could respond to the needs of a diverse user community. We present a prototype of a system used for a virtual laboratory in computational biology.

We introduce ad hoc grids as a hierarchy of mobile devices with different computing and communication capabilities. An ad hoc grid allows a group of individuals to accomplish a mission, often in a hostile environment; examples of applications of ad hoc grids are disaster management, wild-fire prevention, and peacekeeping operations. We are concerned with the interplay between computing and communication in the power-constrained environment of an ad hoc grid.

The simulation and modeling of complex physical systems often involves many components because (i) the physical system itself has components of differing natures, (ii) parallel computing strategies require many (somewhat independent) components, and (iii) existing simulation software applies only to simpler geometrical shapes and physical situations. We discuss how agent based networks are applied to such multi-component applications. The network agents are used to (a) control the execution of existing solvers on sub-components, (b) mediate between sub-components, and (c) coordinate the execution of the ensemble. This paper focuses on partial differential equation (PDE) models as an instance of the approach and describes the implementation of networks using the PELLPACK problem solving environment for PDEs and the Bond system for agent based computing. Copyright © 2000 John Wiley & Sons, Ltd.

In this paper we present a security framework for Bond, a message-oriented distributed object middleware for network computing. Bond Security Framework, BSF, allows developers to exercise performance-security tradeoffs and use the security model best suited for a specific application and for a given environment. BSF consists of an extensible core and a set of well-defined security interfaces. Any Bond object can become a secure object when extended with a dynamic property called bondSecurityContext.