Search

In toy environments like video games, a reinforcement learning agent is deployed and operates within the same state space in which it was trained. However, in robotics applications such as industrial systems or autonomous vehicles, this cannot be guaranteed. A robot can be pushed out of its training space by some unforeseen perturbation, which may cause it to go into an unknown state from which it has not been trained to move towards its goal. While most prior work in the area of RL safety focuses on ensuring safety in the training phase, this paper focuses on ensuring the safe deployment of a robot that has already been trained to operate within a safe space. This work defines a condition on the state and action spaces, that if satisfied, guarantees the robot's recovery to safety independently. We also propose a strategy and design that facilitate this recovery within a finite number of steps after perturbation. This is implemented and tested against a standard RL model, and the results indicate a much-improved performance.

With increased requirements of processing power in many computing applications, and the consequent huge increases in resource costs including electrical power consumption, there is a push to find alternative computing architectures that achieve high speeds but require less resources. One direction that shows promise is brain-inspired neuromorphic computing; it has thus become an active area of research. Much of the current work in neuromorphic computing targets its applicability in solving machine learning and pattern recognition problems. We push this boundary to show the possibility of using neuromorphic architectures to solve matrix multiplication, a core mathematical task pertinent to a host of engineering problems. We use a neuromorphic simulator to present a solution to matrix multiplication that demonstrates a significant performance improvement, both in terms of processing time and intermediate storage over existing approaches. The basic principle of proposed method is operating on a single intermediate transformation matrix that is flattened into an ensemble array represented as a neural node that is operated on. This is an optimisation over memory and time over the existing approach which uses two sparse intermediate matrices. Theoretical and experimental analysis indicates the importance of the proper choice of neuromorphic parameters, specifically the number of neurons and radius of the ensemble along with the dimensions and elements in input matrices has on accuracy, time and memory. We indicate an approach to compute the proper choice of neuromorphic resources given relevant benchmark constraints.

This is a consolidated look at computational techniques for sustainability, and their limits and possibilities. Sustainability is already well established as a concern and a topic of study and practice, given the alarming increase of environmental degradation, pollution, and other adverse effects of industrialization and urbanization. Computational sustainability, which focuses on the use of effective computational models and computational approaches to help achieve the goal of sustainability, has attracted interest from computer science researchers worldwide. We review recent work on computational techniques applied to a range of domains related to sustainability, from bio-surveillance to poverty mapping, from renewable energy production forecasting to crop disease monitoring, and from agent-based modeling to stochastic network design. In sustainable computing, we discuss some directions that have recently been explored. Finally, we analyze research directions that could be explored in the future to achieve the goal of long-term environmental sustainability.

The problem of distributed election majority-winner monitoring for checkpoint-based protocols has been of interest for some time now, and the approach that most of the major algorithms take on this is to employ a center-based pull from all voting sites in conjunction with a count tracker, to continually monitor the information about the incoming voter stream. This paper presents an alternative solution to this problem, using site-based voter information push to the center, with reduced communication complexity, using a unique sampling technique. This technique is utilized to come up with a winner or an approximate winner with very high probability for Approval-based and Scoring-based rules. The site-push algorithm introduced here has been validated with correctness results, and has also been analyzed for different scenarios to see the improvement in the number of messages initiated, in comparison to the center-based pull algorithm considered for a checkpoint-based protocol. The effect of this algorithm is to reduce the communication complexity by a factor of $(1+\frac{\log {k}}{\log {\frac{n}{k}}})$ in comparison to the center-pull algorithm (where $k$ is the number of voting sites, and $n$ the number of voters), which leads to a large reduction in communication complexity in democratic election scenarios. Also, to analyze other scenarios, simulation results are presented to show how the algorithm fares under randomized voter assignment, with variations in the value of total voters, total sites and the approximation parameter $\epsilon$ .

Neuromorphic computing is looked at as one of the promising alternatives to the traditional von Neumann architecture. In this paper, we consider the problem of doing arithmetic on neuromorphic systems and propose an architecture for doing IEEE 754 compliant addition on a neuromorphic system. A novel encoding scheme is also proposed for reducing the inter-neural ensemble error. The complex task of floating point addition is divided into sub-tasks such as exponent alignment, mantissa addition and overflow-underflow handling. We use a cascaded approach to add the two mantissas of the given floating-point numbers and then apply our encoding scheme to reduce the error produced in this approach. Overflow and underflow are handled by approximating on XOR logic. Implementation of sub-components like right shifter and multiplexer are also specified.

Humans can easily parse and find answers to complex queries such as "What was the capital of the country of the discoverer of the element which has atomic number 1?" by breaking them up into small pieces, querying these appropriately, and assembling a final answer. However, contemporary search engines lack such capability and fail to handle even slightly complex queries. Search engines process queries by identifying keywords and searching against them in knowledge bases or indexed web pages. The results are, therefore, dependent on the keywords and how well the search engine handles them. In our work, we propose a three-step approach called parsing, tree generation, and querying (PTGQ) for effective searching of larger and more expressive queries of potentially unbounded complexity. PTGQ parses a complex query and constructs a query tree where each node represents a simple query. It then processes the complex query by recursively querying a back-end search engine, going over the corresponding query tree in postorder. Using PTGQ makes sure that the search engine always handles a simpler query containing very few keywords. Results demonstrate that PTGQ can handle queries of much higher complexity than standalone search engines.

The "Thinking, Fast and Slow" paradigm of Kahneman proposes that we use two different styles of thinking -- a fast and intuitive System 1 for certain tasks, along with a slower but more analytical System 2 for others. While the idea of using this two-system style of thinking is gaining popularity in AI and robotics, our work considers how to interleave the two styles of decision-making, i.e., how System 1 and System 2 should be used together. For this, we propose a novel and general framework which includes a new System 0 to oversee Systems 1 and 2. At every point when a decision needs to be made, System 0 evaluates the situation and quickly hands over the decision-making process to either System 1 or System 2. We evaluate such a framework on a modified version of the classic Pac-Man game, with an already-trained RL algorithm for System 1, a Monte-Carlo tree search for System 2, and several different possible strategies for System 0. As expected, arbitrary switches between Systems 1 and 2 do not work, but certain strategies do well. With System 0, an agent is able to perform better than one that uses only System 1 or System 2., 7 pages, 11 figures; typos corrected, references added

Bundling is a technique e-commerce companies have adopted from traditional retail stores to increase the average order size. It has been observed that bargaining helps increase customer satisfaction while increasing the average order revenue for retailers. We propose a mathematical framework to incorporate bargaining capabilities with the product bundles provided by e-commerce websites. Our method creates a virtual agent that uses the modular Bidding-Opponent-Acceptance model for its bargaining strategy and the Thomas-Kilmann conflict mode instrument to model buyer behavior. We incorporate bargaining capabilities with bundles in an e-commerce system by using a negotiation agent that uses business logic for better strategy. It uses real-time data generated during a negotiation session, since the buyer behavior during a negotiation is crucial. No requirement exists for data from past negotiation sessions of the buyer, which removes bias as well as allowing for rapid changes to buyer behavior. The agent behavior can be altered by various hyperparameters. Our model provides utility metrics to measure buyer and agent satisfaction. Our results show that the agent successfully negotiates with humans from diverse backgrounds.

A problem currently faced by autonomous systems, learning agents, multi-agent systems and other cognitive systems is the extreme susceptibility to bad information and malicious misdirection. In this paper, we propose a strategy to identify trusted agents in a multi-agent system, so that a system can eliminate the knowledge provided by malicious or unknown agents. The strategy is covered in three phases by each agent; firstly sharing information with other agents in the system, secondly calculating the conclusion based on the majority of information received from others and lastly identifying the trusted, malicious and unknown agents in the system. The theoretical analysis shows that the trusted agent strategy (TAS) is capable of identifying malicious or compromised agents if a system has less than 25% of malicious agents. This is further improved in the advanced-TAS which can handle up to 40% malicious agents. The experimental results show that trusted and malicious agents can be identified with high accuracy.

In hybrid cloud computing, cloud users have the ability to procure resources from multiple cloud vendors, and furthermore also the option of selecting different combinations of resources. The problem of procuring a single resource from one of many cloud vendors can be modeled as a standard winner determination problem, and there are mechanisms for single item resource procurement given different QoS and pricing parameters. There however is no compatible approach that would allow cloud users to procure arbitrary bundles of resources from cloud vendors. We design the cloud - $\mathcal {CABOB}$ algorithm to solve the multiple resource procurement problem in hybrid clouds. Cloud users submit their requirements, and in turn vendors submit bids containing price, QoS and their offered sets of resources. The approach is scalable, which is necessary given that there are a large number of cloud vendors, with more continually appearing. We perform experiments for procurement cost and scalability efficacy on the cloud - $\mathcal {CABOB}$ algorithm using various standard distribution benchmarks like random, uniform, decay and CATS. Simulations using our approach with prices procured from several cloud vendors’ datasets show its effectiveness at multiple resource procurement.

Much of today's business and research is driven by data and the analysis of large volumes of it. Big-data analytics has become a key phrase in computing. When it comes to business, it plays a crucial role in strategizing and decision making, and it provides insight on performance and usage. Areas of research are driven by the need to create new models to better understand and analyze data.

During mass evacuations, many psychological and physical factors are responsible for stampedes and other life threatening situations. Quantitative and qualitative analyses of these factors are of high importance while devising optimal strategies for evacuations. In this paper, we present an agent-based model that considers psychological and physical factors that cause panic in such situations. We have also simulated some simple evacuation scenarios and presented a method to identify possible bottlenecks and shortcomings in the environments during emergency evacuations. Our method also helps in evaluation and analysis of different evacuation strategies. To enable this analysis, we have used a rule-based roadmap approach, where critical nodes in the environment are identified by the evacuation planner and each node has a special rule according to the strategy of the planner. We evaluate different strategies on parameters, such as evacuation time and physical discomfort caused to the agents.

The rise in the penetration of the internet across the world has led to a rapid increase in the consumption of energy at the data centers established by leading cloud data service providers. High power consumption by these data centers [DCs] leads to high operational costs and high carbon emissions into the environment. From a sustainability point of view, the ultimate goal is to maximize the productivity and efficiency of these data centers while keeping greenhouse gas emissions to the minimum and maximize data center productivity. This goal can be achieved by better resource utilization and replacing carbon-intensive approaches of energy production with green sources of energy. Due to the limited intermittent availability of renewable sources of energy, the ideal ‘Green’ design for the DCs, should incorporate inter-operability with both renewable and non-renewable sources of energy. In this paper, we propose a ren-aware scheduler to schedule computational workload by prioritizing their execution within the duration of green energy availability on the basis of the predicted hourly green energy and workload data of DCs. Our results demonstrate that our ren-aware scheduler can increase the green energy consumption by 51% compared to the conventional randomized scheduler that distributes load without considering green energy and load. It can also reduce the total energy consumption by 25% by putting the DCs to sleep during their idle time, as it saves 4.5 times more idle energy than the randomized scheduler. Additionally, the results also demonstrate how the role of time zones of the DCs and the duration of green energy availability in them is pivotal in our ren-aware scheduler's performance.

In IT service delivery and support, the cloud paradigm has introduced the problem of resource over-provisioning through rapid automation (or orchestration) of manual IT operations. Due to the elastic nature of cloud computing, this shortcoming ends up significantly reducing the real benefit, viz., the cost-effectiveness of cloud adoption for Cloud Service Consumers (CSC). Similarly, detecting and eliminating such over-provisioning of cloud resources without affecting the quality of service (QoS) is extremely difficult for Cloud Service Providers (CSPs) since they have visibility only into the state of the IT services (cloud resources) but none into the actual performance of business services. In this paper, we propose Tier-centric Business Impact and Cost Analysis (T-BICA), a tier-centric optimal resource allocation algorithm, to address the problem of rapid provisioning of IT resources in modern enterprise cloud environments, through extensive data gathering and performance analyses of business services in a simulated environment emulating a mature cloud service provider. We have derived improved analytics to address the issues and to accelerate real cloud adoption for large enterprises within the context of meeting (or exceeding) business service level objectives (SLOs) and minimizing the cloud subscription cost (OpEx) for the business. While investigating the problem, we consider the time and the cost of delivering business service in medium- to large-size enterprise environments, quantifying the negative impact of IT resource over-provisioning (due to highly mature IT services centric orchestration capabilities) on the business, and indicate how the suggested cloud analytics could assist in reducing total cost of ownership (TCO) of the business service. From our analysis of the test data, we have observed that our suggested approach and analytic reduces the cost of delivering business services by 65.19 percent, and improves the performance (total time to deliver) by 74.18 percent when compared to the existing modern cloud management and resource provisioning approach. Using T-BICA dramatically reduces upfront costs (CapEx) for CSPs (from the capacity procurement and management points of view) through efficient on-demand resource de-provisioning, without affecting business SLOs and IT service level agreements (SLAs). The improved dynamic allocation of resources also makes for better efficiency of utilization, which in turn has desirable consequences for sustainability, and makes this an approach for “Green” IT.

This paper presents an integrated model that approximates pedestrian behavior in case of a fire emergency and its consequences. We have modeled a confined fire with a variable spread rate, based on the existing literature pertaining to the field. The fire has both psychological and physical impacts on the state of the agents. The model also incorporates clustering behavior in agents, which slows down the evacuation. The model helps recognize bottlenecks and compares the evacuation efficiency by comparing casualties across different scenarios. Simulation results are given as illustrations and give qualitative insights into the risks and likely problems in specific fire scenarios.

Drug-drug interaction causes potential impact on patients when a second drug is administered during the duration of action of the first. It may result in the delay or decrease in the absorption of rate of drugs or enhance their absorption. This also in turn may affect the action of drugs or induce adverse effects on patients. There exists a need to study the drug-drug interactions, and their potential effects on the human system, including for drugs not yet approved. This paper proposes using eight features (substructure, targets, transporters, enzymes, pathways, indications, side-effect and off-side-effect, obtained from five different databases: PubChem, Drugbank, KEGG, SIDER, Offsides, and a similarity-based ensemble prediction model to identify the potential drug-drug interactions. The ensemble model uses the Jaccards coefficient method for identifying similarity measures between drugs. This similarity indices are given to a neighbor recommender method and random walk method for the base prediction of drug-drug interaction. This predictive model is improved by an ensemble modrisk and may be of little clinical significance. It can take years to clinically check DDIs for every pair of drugs. Sometimes DDIs may not get detected in clinical trials. Moreover, it may take many years to check DDIs of all known drugs with a newly discovered one before it is introduced to the market. Hence, there exists the need for efficient DDI-checking with fewer time-consuming, expensive, and risky clinical trials.

Virtualization of resources on the cloud offers a scalable means of consuming services beyond the capabilities of small systems. In a cloud that offers infrastructure such as processor, memory, hard disk, etc., a coalition of virtual machines formed by grouping two or more may be needed. Economical management of cloud resources needs allocation strategies with minimum wastage, while configuring services ahead of actual requests. We propose a resource allocation mechanism for machines on the cloud, based on the principles of coalition formation and the uncertainty principle of game theory. We compare the results of applying this mechanism with existing resource allocation methods that have been deployed on the cloud. We also show that this method of resource allocation by coalition-formation of the machines on the cloud leads not only to better resource utilization but also higher request satisfaction.

Vaccine development is a laborious and time-consuming process and can benefit from statistical machine learning techniques, which can produce general outcomes based on the patterns observed in the limited available empirical data. In this paper, we show how limited gene expression data from a small sample of subjects can be used to predict the outcomes of malaria vaccine. In addition, we also draw inferences from the gene expression data, with over 22000 columns (or features), by visualizing the data, and reduce the data dimensions based on this inference for efficient model training. Our methods are general and reliable and can be extended to vaccines developed against any pathogen. Given the gene expression data from a sample of subjects administered with a novel vaccine, our methods can be used to test the outcome of that vaccine, without the need for empirical observations on a larger population. By carefully tuning the available data and the machine learning models, we are able to achieve greater than 98% accuracy, with sensitivity and specificity of 0.93 and 1 respectively, in predicting the outcomes of the malaria vaccine in developing immunogenicity against the malaria pathogen.

Due to its widespread practical applications to image database management, semantic image retrieval has received a lot of attention in past decades. In particular, relevance feedback-based methods have been a popular approach toward bridging the gap between low-level features and high-level semantic concepts. Nature-inspired algorithms have shown a lot of potential to solve problems which are complex and require discovering of patterns under changing environment. However, lack of sound mathematical foundations has been considered a drawback toward better analysis of these algorithms. In this chapter, we propose a novel general topological model for semantic image retrieval using relevance feedback. We use point-set topology to develop mathematical constructs for modeling the semantic retrieval. In particular, we develop an image retrieval algorithm based on the ant sleeping model and extend the topological model to analyze it. Through experiments we show that our algorithm performs well in indexing an image database for relevance feedback. With our indexing procedure, the average response time to access image results from a storage device is lower when compared to vector quantization techniques. We also evaluate our algorithm, theoretically and empirically, against PicSOM (a CBIR system based on relevance feedback). Our ASM-based technique shows a very efficient retrieval performance using relevance feedback.

High sparsity and the problem of overspecialization are challenges faced by collaborative filtering (CF) algorithms in recommender systems. In this paper, we design an approach that efficiently tackles the above problems. We address the first issue of high sparsity in CF by modifying the popular parallel seeding technique proposed by Bahmani et al. for use in a spherical setting, for the clustering of users. Experimental evaluations on highly sparse real world datasets demonstrate the better performance of our algorithm than existing Spherical K-Means algorithms. Contrary to the common belief that users are only interested in items similar to those of their previous liking, it has been well established that including serendipitous recommendations improves their satisfaction. Thus, to tackle the second problem of overspecialization, we effectuate serendipity in movie recommender systems with an end-to-end algorithm, Serendipitous Clustering for Collaborative Filtering (SC-CF) that considers diversity, unexpectedness and relevance. SC-CF takes advantage of carefully generated user profiles to then assign users to a "serendipitous cluster." The overall clustering process leverages these user profiles and improves recommendation quality through serendipity. A series of experiments on Serendipity 2018 (part of Movielens) dataset built on real user feedback has shown that SC-CF outperforms the existing popular recommendation methods like K-Means and deep learning based CF.

Permanent memory, and biases that humans have such as confirmation bias, are not modeled in some of the leading strategies in Iterated Prisoner's Dilemma. This reduces their effectiveness at modeling actual human behaviors. As a solution to this problem we have proposed a framework to model egoistic agents using trust. This framework accounts for such flaws in human behavior, and provides a more accurate representation of human nature by modeling confirmation bias through egoism and trust. The agents of this model divide their neighboring agents into two sets, I and Non-I, and the actions taken against a neighbor are influenced by the set it belongs to. This framework can be used to modify existing strategies such as Tit-For-Tat to make them more realistic. Using this framework, we have also proposed a trust based strategy and evaluated it. Our strategy does better than the other strategies in pairwise comparisons, and also when there is a combination of different agents in the environment.

As military, academic, and commercial computing systems evolve from autonomous entities that deliver computing products into network centric enterprise systems that deliver computing as a service, opportunities emerge to consolidate computing resources, software, and information through cloud computing. Along with these opportunities come challenges, particularly to service providers and operations centers that struggle to monitor and manage quality of service (QoS) for these services in order to meet customer service commitments. Traditional approaches fall short in addressing these challenges because they examine QoS from a limited perspective rather than from a system-of-systems (SoS) perspective applicable to a net-centric enterprise system in which any user from any location can share computing resources at any time. This paper presents a SoS approach to enable QoS monitoring, management, and response for enterprise systems that deliver computing as a service through a cloud computing environment. A concrete example is provided for application of this new SoS approach to a real-world scenario (viz., distributed denial of service). Simulated results confirm the efficacy of the approach.

Achieving a balance of supply and demand in a multi-agent system with many individual self-interested and rational agents that act as suppliers and consumers is a natural problem in a variety of real-life domains---smart power grids, data centers, and others. In this paper, we address the profit-maximization problem for a group of distributed supplier and consumer agents, with no inter-agent communication. We simulate a scenario of a market with $S$ suppliers and $C$ consumers such that at every instant, each supplier agent supplies a certain quantity and simultaneously, each consumer agent consumes a certain quantity. The information about the total amount supplied and consumed is only kept with the center. The proposed algorithm is a combination of the classical additive-increase multiplicative-decrease (AIMD) algorithm in conjunction with a probabilistic rule for the agents to respond to a capacity signal. This leads to a nonhomogeneous Markov chain and we show almost sure convergence of this chain to the social optimum, for our market of distributed supplier and consumer agents. Employing this AIMD-type algorithm, the center sends a feedback message to the agents in the supplier side if there is a scenario of excess supply, or to the consumer agents if there is excess consumption. Each agent has a concave utility function whose derivative tends to 0 when an optimum quantity is supplied/consumed. Hence when social convergence is reached, each agent supplies or consumes a quantity which leads to its individual maximum profit, without the need of any communication. So eventually, each agent supplies or consumes a quantity which leads to its individual maximum profit, without communicating with any other agents. Our simulations show the efficacy of this approach., Comment: 6 pages, 12 figures, IEEE International Conference on Systems, Man and Cybernetics

We present ASD (Action, Sequence, and Divide), a new framework for Hierarchical Reinforcement Learning (HRL). Present HRL methods construct the task hierarchies but fail to avoid exploration when tasks are to be performed in a particular sequence, resulting in the agent needlessly exploring all permutations of the tasks. When the task hierarchies are used as an ASD framework, the RL agent encounters better constraints, preventing it from pursuing policies that are not valid, thus enabling the agent to achieve the optimal policy faster. The hierarchies created using the methods explained in this paper can be used to solve new episodes of the same environment, as well as similar instances of the problem. The hierarchies generated with an ASD framework can be used to establish an ordering of tasks. The objective is to not only to complete the tasks but also give the agent insights into the sequence of tasks that need to be performed in order to correctly solve a problem. We present an algorithm to generate the hierarchies as an ASD framework. The algorithm has been evaluated on some of the standard RL domains, namely, Taxi and Wargus, and is found to give correct results.

Neural Machine Translation (NMT) is a new technique for machine translation that has led to remarkable improvements compared to rule-based and statistical machine translation (SMT) techniques, by overcoming many of the weaknesses in the conventional techniques. We study and apply NMT techniques to create a system with multiple models which we then apply for six Indian language pairs. We compare the performances of our NMT models with our system using automatic evaluation metrics such as UNK Count, METEOR, F-Measure, and BLEU. We find that NMT techniques are very effective for machine translations of Indian language pairs. We then demonstrate that we can achieve good accuracy even using a shallow network; on comparing the performance of Google Translate on our test dataset, our best model outperformed Google Translate by a margin of 17 BLEU points on Urdu-Hindi, 29 BLEU points on Punjabi-Hindi, and 30 BLEU points on Gujarati-Hindi translations.

An extensible, decentralized network of self-governing objects connected to a shared but variable power supply is a realistic problem domain for certain demand-management problems arising in the context of futuristic systems connected to smart power grids. An algorithmic framework of such a network is presented and discussed in this paper. Each object of the network has a power demand and a priority, and it is interconnected and able to exchange information with all other objects in the system. As each object shares information (its power demand and priority) with the other objects, the system as a whole exhibits self-governance, and is able to be managed without a centralized controller. Our model, which we term as decentralized power distribution (DPD) for such a network, also allows us to formulate algorithms for distributing available power among objects, along with analyses of correctness and performance.

We propose a novel distributed approach for the problem of scheduling a fleet of autonomous vehicles. Our distributed system avoids a single point of failure, is scalable, fault tolerant and robust. We describe an agent-based distributed system to conduct a set of procurement auctions. The vehicles are the “sellers” and the passengers are the “buyers” in the auction. Each vehicle bids for the passengers with a bid value which is an inverse function of the time it would take the vehicle to reach the passenger. In our agent based architecture, the various software agents reside on different systems and we describe distributed algorithms for their communication. We have performed simulations of a vehicle fleet in two different locations (Bangalore, India and Tyson's Corner, Virginia, USA) and compare the maximum and average waiting times of the passenger of our algorithm with a FIFO algorithm. We also compute the ratio of the time in which the vehicle is servicing a passenger to the total time to compute the fuel wastage. The results show that our system improves the maximum and average waiting times of the passengers, as well as the fuel costs for the vehicle fleet. Furthermore, we show that such a distributed system reduces the time it would take on average to respond to customer requests, as compared to a system which is not distributed.

Scheduling of tasks on a multi-machine system to reduce the makespan, while satisfying the precedence constraints between the tasks, is known to be an NP-hard problem. We propose a new formulation and show that energy-aware scheduling is a generalization of the minimum makespan scheduling problem. Taking the system graph and program graph as inputs, we propose three different algorithms for energy-aware scheduling, each of them having its own strengths and limitations. The first is a genetic algorithm (Plain GA) that searches for an energy reducing schedule. The second (CA+GA) uses cellular automata (CA) to generate low energy schedules, while using a genetic algorithm (GA) to find good rules for the CA. The third (EAH) is a heuristic which gives preference to high-efficiency machines in allocation. We have tested our algorithms on well-known program graphs and compared our results with other state-of-the-art scheduling algorithms, which confirms the efficacy of our approach. Our work also gives insight into the time-energy trade-offs in scheduling.

The cost of electricity contributes significantly to the operating expense incurred in hosting cloud services. It is necessary to consider this cost while charging the consumers for their service utilization. In this work, we arrive at a metering mechanism for cloud services, in which the price of a cloud service tracks the variable input cost of electricity from a smart grid. The power-aware cloud metering developed here is a dynamic pricing and billing model where tariff for a cloud service is varied in accordance with the input electricity cost. We arrive at a model for power consumption of virtual machines hosted on the cloud infrastructure. This power consumption model is used in calculating the cost of operation of the service. A cloud instance leased by a consumer is billed based on the cost of operation obtained, and its resource utilization. Experimental results validate the approach presented.

Increasing energy costs of large-scale server systems have led to a demand for innovative methods for optimizing resource utilization in these systems. Such methods aim to reduce server energy consumption, cooling requirements, carbon footprint, and so on, thereby leading to improved holistic sustainability of the overall server infrastructure. At the core of many of these methods lie reliable workload-prediction techniques that guide in identifying servers, time intervals, and other parameters that are needed for building sustainability solutions based on techniques like virtualization and server consolidation for server systems. Many workload prediction methods have been proposed in the recent paper, but unfortunately they do not deal adequately with the issues that arise specifically in large-scale server systems, viz., extensive nonstationarity of server workloads, and massive online streaming data. In this paper, we fill this gap by proposing two online ensemble learning methods for workload prediction, which address these issues in large-scale server systems. The proposed algorithms are motivated from the weighted majority and simulatable experts approaches, which we extend and adapt to the large-scale workload prediction problem. We demonstrate the effectiveness of our algorithms using real and synthetic data sets, and show that using the proposed algorithms, the workloads of 91% of servers in a real data center can be predicted with accuracy > 89%, whereas using baseline approaches, the workloads of only 13%–24% of the servers can be predicted with similar accuracy.

We present a cloud resource procurement approach which not only automates the selection of an appropriate cloud vendor but also implements dynamic pricing. Three possible mechanisms are suggested for cloud resource procurement: cloud-dominant strategy incentive compatible (C-DSIC), cloud-Bayesian incentive compatible (C-BIC), and cloud optimal (C-OPT). C-DSIC is dominant strategy incentive compatible, based on the VCG mechanism, and is a low-bid Vickrey auction. C-BIC is Bayesian incentive compatible, which achieves budget balance. C-BIC does not satisfy individual rationality. In C-DSIC and C-BIC, the cloud vendor who charges the lowest cost per unit QoS is declared the winner. In C-OPT, the cloud vendor with the least virtual cost is declared the winner. C-OPT overcomes the limitations of both C-DSIC and C-BIC. C-OPT is not only Bayesian incentive compatible, but also individually rational. Our experiments indicate that the resource procurement cost decreases with increase in number of cloud vendors irrespective of the mechanisms. We also propose a procurement module for a cloud broker which can implement C-DSIC, C-BIC, or C--OPT to perform resource procurement in a cloud computing context. A cloud broker with such a procurement module enables users to automate the choice of a cloud vendor among many with diverse offerings, and is also an essential first step toward implementing dynamic pricing in the cloud.

This paper presents the best practices to carry out the verification and validation (V&V) for a safety critical embedded system, part of a larger system-of-systems. The paper talks about the effectiveness of this strategy from performance and time schedule requirement of a project. The best practices employed for the V &Vis a modification of the conventional V&V approach. The proposed approach is iterative which introduces new testing methodologies apart from the conventional testing methodologies, an effective way of implementing the phases of the V&V and also analyzing the V&V results. The new testing methodologies include the random and non-real time testing apart from the static and dynamic tests. The process phases are logically carried out in parallel and credit of the results of the different phases are taken to ensure that the embedded system that goes for the field testing is bug free. The paper also demonstrates the iterative qualities of the process where the iterations successively find faults in the embedded system and executing the process within a stipulated time frame, thus maintaining the required reliability of the system. This approach is implemented in the most critical applications —-aerospace application where safety of the system cannot be compromised. The approach used a fixed number of iterations which is set to4in this application, with each iteration adding to the reliability and safety of the embedded system. Data collected and results observed are compared with a conventional approach for the same application and it is demonstrated that the strategy proposed reduces the time taken by 50% as compared to a conventional process that attains the same reliability as required in the stipulated time.

Cloud computing is not only a popular paradigm for services offered over the Internet, but has also captured the interest of both academia and industry.

The rapid spread of fake news is a serious problem calling for AI solutions. We employ a deep learning based automated detector through a three level hierarchical attention network (3HAN) for fast, accurate detection of fake news. 3HAN has three levels, one each for words, sentences, and the headline, and constructs a news vector: an effective representation of an input news article, by processing an article in an hierarchical bottom-up manner. The headline is known to be a distinguishing feature of fake news, and furthermore, relatively few words and sentences in an article are more important than the rest. 3HAN gives a differential importance to parts of an article, on account of its three layers of attention. By experiments on a large real-world data set, we observe the effectiveness of 3HAN with an accuracy of 96.77%. Unlike some other deep learning models, 3HAN provides an understandable output through the attention weights given to different parts of an article, which can be visualized through a heatmap to enable further manual fact checking.

We present EEQuest, an application that extracts events from text using natural language processing (NLP) and supervised machine-learning techniques, and provides a system to query events extracted from a text corpus. We provide a use case for the application wherein we extract business-related events from news articles. The extracted events are then categorized based on the business organization/company that they are related to. Finally, the events are added to a knowledge base using which a query system is built. The system can be used to display events related to a particular organization or a group of organizations. Although we are using the system to extract business-related events, the event extraction mechanism can be used in a more general sense with any available textual data, to extract any kind of events that have a structure that can answer the question: Who did what, when and where?

A cellular automata (CA) configuration is constructed that exhibits emergent failover. The configuration is based on standard Game of Life rules. Gliders and glider-guns form the core messaging structure in the configuration. The blinker is represented as the basic computational unit, and it is shown how it can be recreated in case of a failure. Stateless failover using primary-backup mechanism is demonstrated. The details of the CA components used in the configuration and its working are described, and a simulation of the complete configuration is also presented., Comment: 16 pages, 15 figures and associated video at http://dl.dropbox.com/u/7553694/failover_demo.avi and simulation at http://dl.dropbox.com/u/7553694/failover_simulation.jar

A framework is proposed for an electrical power microgrid, such as for a colony or small township of homes that generate electrical power from solar energy and use it directly when possible, and via stored battery power at other times. The situation is described as a demand and supply problem in a multi-agent system with many consumers and suppliers and no explicit communication or coordination among the agents. Such a demand and supply problem is modeled as a Potluck Problem, a generalization of the Santa Fe Bar Problem. Power produced by PV panels and batteries may be used in the local market, in addition to being consumed locally. The proposed microgrid system model is able to determine the optimum operation of a solar-powered microgrid with respect to load demand, environmental requirements, PV panel and battery capacities. The results indicate the effect of various such parameters on the performance of these micro-grids. This paper also analyzes and proposes, based on auction theory, the most efficient and competing pricing mechanism in the proposed microgrid system model. Two important market bidding techniques, single bidding and discriminatory bidding, are considered. The microgrid is made to participate in the bidding process to serve the consumers at a reduced price and to provide better revenues. The viability of the model proposed is illustrated with analyses using realistic assumptions and published historical data.

Astronomical datasets are typically very large, and manually classifying the data in them is effectively impossible. We use machine learning algorithms to provide classifications (as stars, quasars and galaxies) for more than one billion objects given photometrically in the Third Data Release of the Sloan Digital Sky Survey (SDSS-III). We have used k NN, SVM and random forest algorithms in a distributed environment over the cloud to classify 1,183,850,913 unclassified photometric objects present in the SDSSIII catalog. This catalog contains photometric data for all objects viewed through a telescope and spectroscopic data for a small part of these. Although it is possible to classify all the objects using spectroscopic data, it is impractical to obtain such data for each one of them. To classify such a big dataset on a single machine would be impractically slow, so we have used the Spark cluster computing framework to implement a distributed computing environment over the cloud. We found that writing results (dozens of gigabytes) to the cloud storage is very slow while using k NN. Though writing the results with SVM is faster as it is done in parallel, its accuracy is only around 87%, due to lack of a kernel implementation of it in Spark. We then used the random forest algorithm to classify the entire set of 1,183,850,913 objects with an accuracy of 94% in about 17 hours of processing time. The result set is significant as even collecting spectroscopic data for these many objects would take decades, and our classifications can help astronomers and astrophysicists carry out further studies.

The problem of attaining energy efficiency in distributed systems is of importance, but a general, non-domain-specific theory of energy-minimal scheduling is far from developed. In this paper, we classify the problems of energy-minimal scheduling and present theoretical foundations of the same. We derive results concerning energy-minimal scheduling of independent jobs in a distributed system with functionally similar machines with different working and idle power ratings. The machines considered in our system can have identical as well as different speeds. If the jobs can be divided into arbitrary parts, we show that the minimum-energy schedule can be generated in linear time and give exact scheduling algorithms. For the cases where jobs are non-divisible, we prove that the scheduling problems are NP-hard and also give approximation algorithms for the same along with their bounds., Comment: 41 pages

It is known that energy-minimal scheduling is strictly harder than minimal-makespan scheduling, but it is not well understood. This paper is a study of the problems of scheduling to minimize the energy consumed by a system of dissimilar machines (with no restriction on the makespan). The different system types considered are for machines with identical speeds but different power ratings, and for machines with different speeds and power ratings but no idle power consumption. The results give insights for further studies classifying various types of energy-minimal scheduling problems, and for deriving algorithms for exact or approximate solutions for such problems.

International audience; Chap.8 SCHEDULING AND RESOURCE ALLOCATION - 8.1 Introduction: Energy-Aware Scheduling 8.2 Use of Linear Programming in Energy-Aware Scheduling 8.2.1 Finding the Optimal Solution Using a Linear Program 8.2.2 Benefits and Limitations of LP 8.3 Heuristics in Large Instances 8.3.1 Energy-Aware Greedy Algorithms 8.3.2 Vector Packing 8.3.3 Improving Fast Algorithms 8.4 Comparing Allocation Heuristics for Energy-Aware Scheduling 8.4.1 Problem Formulation 8.4.2 Allocation Heuristics 8.4.3 Results 8.5 Energy-Aware Task Allocation in Mobile Environments 8.5.1 Reference Architecture 8.5.2 Task Allocation Strategy 8.5.3 Task Allocation Algorithm 8.5.4 Performance Results 8.6 An Energy-Aware Scheduling Strategy for Allocating Computational Tasks in a Fully Decentralized Way 8.6.1 Decentralized Resources in Cloud: Overview 8.6.2 Cooperative Scheduling Anti-Load Balancing Algorithm for Cloud (CSAAC) 8.6.3 Simulation Results 8.6.4 Evaluation 8.7 Cost-Aware Scheduling with Smart Grids 8.7.1 Cost-Aware Scheduling 8.7.2 Cost-Aware Scheduling Using DE 8.7.3 Comparison of DE with Other Approaches 8.8 Heterogeneity, Cooling, DVFS, and Migration 8.8.1 Lever Interactions 8.8.2 Infrastructures 8.8.3 Resource Allocation as a Whole 8.9 Conclusions - References

Apache Hadoop is an open source framework for processing massive amount of data in a distributed environment. Hadoop services use a polling mechanism for event notifications. In this paper, we propose a distributed notification system for Hadoop based on the Publish-Subscribe model. Such a notification system can be used for message-passing among Hadoop services. It can also be used to chain multiple MapReduce jobs based on events occuring in a Hadoop cluster. This results in reduced cluster load and network bandwidth requirement. We have used two popular Publish-Subscribe-based messaging systems—Apache ActiveMQ and Apache Kafka—for implementation. Lastly, we have executed performance tests on both these messaging systems to monitor time taken for message delivery and reception.

Predictive workload analytics for server systems has been the focus of recent research in energy-aware computing, with many algorithmic and architectural techniques proposed to analyze, predict, and optimize server workloads in order to build energy-aware IT systems. These techniques, though effective in optimizing server workloads, often ignore the green-ness aspect 'in' the technique itself and incur heavy computational costs in their operations. In this paper, we propose a meta-learning based architecture for building server workload prediction mechanism using which the computational cost of holistic predictive workload analytics can be optimized, and green-ness 'in' the analytics can be achieved. We also present an algorithmic optimization of the proposed meta-learning architecture for handling concept drift in server workloads, thereby also achieving improved greenness 'by' the analytics. Our experiments show that the proposed meta-learning based architecture substantially reduces the total computational cost of workload prediction mechanism, with a minor decrease of 0.5--1.3% in the accuracy, and the proposed algorithmic optimization significantly improves accuracy of the workload prediction mechanism in concept drift scenarios by 8.1%.

We have implemented a Hadoop compatible framework that helps in detection of suspected hardware failures in DataNodes within a Hadoop cluster and in signalling the nodes accordingly. Based on the status of the various hardware components of the DataNodes, the master node signals the DataNode so that appropriate actions can be taken. In a Hadoop cluster, the knowledge of the network is important for the master node to configure itself to meet varying needs of the DataNodes. It is important to keep track of the current topology of the Hadoop network and track the status of the network services in Hadoop. We have also added the functionality to discover the network topology in a Hadoop cluster. This discovered topology information will be useful in performing load balancing in the network, or in making intelligent decisions in data replication in case of node failures.

Growing scale of server infrastructure in large datacenters has led to an increased need for effective server workload prediction mechanisms. Two main challenges faced in server workload prediction task are lack of large-scale training data and changes in the underlying distribution of server workloads in events like change in dominant applications of servers or change in allocation of servers, etc. In this work, we propose an online server workload prediction approach based on ensemble learning which addresses these issues. We evaluate the proposed approach using real dataset of an enterprise data center and a synthetic dataset. Experimental results reveal that the proposed approach achieves accuracy of 87.8% on real dataset and 88.8% on synthetic dataset.

We present a virtualized setup of a Hadoop cluster that provides greater computing capacity with lesser resources, since a virtualized cluster requires fewer physical machines. The master node of the cluster is set up on a physical machine, and slave nodes are set up on virtual machines (VMs) that may be on a common physical machine. Hadoop configured VM images are created by cloning of VMs, which facilitates fast addition and deletion of nodes in the cluster without much overhead. Also, we have configured the Hadoop virtualized cluster to use capacity scheduler instead of the default FIFO scheduler. The capacity scheduler schedules tasks based on the availability of RAM and virtual memory (VMEM) in slave nodes before allocating any job. So instead of queuing up the jobs, they are efficiently allocated on the VMs based on the memory available. Various configuration parameters of Hadoop are analyzed and the virtualized cluster is fine-tuned to ensure best performance and maximum scalability.