We present knowledge-aware iterative retrieval (KAIR), an innovative framework for LLM agents driven by large language models (LLMs) that iteratively refines user queries and filters contextual evidence using dynamically evolving knowledge to facilitate better systems. A key feature of this system is its separation of external information sources from an internal knowledge cache, which is continuously updated to guide query formulation and evidence selection. This design reduces the risk of biased feedback loops and facilitates dynamic, trackable exploration of search paths, balancing the exploration of diverse information with precision through autonomous agent decision-making. Our methodology is assessed using a comprehensive set of open-domain question–answering benchmarks, covering multistep tasks that require synthesizing information from multiple sources, particularly given LLMs’ limitations in explicit reasoning and planning. The results indicate that KAIR not only surpasses single-step baseline models regardless of task complexity but also shows significant advantages over traditional iterative retrieval approaches in challenging tasks by employing precise evidence-based reasoning and enhanced efficiency. Additionally, KAIR supports both competitive and collaborative sharing of updated context, facilitating multiagent extension.

We present COSMIC, a leader-driven multiagent coordination framework that extends the AutoGen GroupChat paradigm to support more stable and context-consistent collaboration among large language model (LLM) agents. COSMIC introduces: 1) a leader agent that decomposes tasks and issues an explicit termination decision; 2) a shared rolling memory buffer that synchronizes recent context across agents; and 3) a speaker selector agent (SSA) that enforces structured turn-taking and mitigates self-invocation and speaker imbalance. Across six coordination-intensive tasks, COSMIC achieves 100% validity under task-specific hard-constraint criteria. On the five tasks where baseline frameworks produced validated outputs, COSMIC improves composite scores by 20.2%, 23.3%, 27.7%, 52.6%, and 46.9% over CrewAI Hierarchical, CrewAI Sequential, AutoGen, LangChain, and OpenAI Swarm, respectively. COSMIC reduces wall-clock time relative to CrewAI Hierarchical, CrewAI Sequential, AutoGen, and LangChain by 57.3%, 50.8%, 61.4%, and 77.6%, while incurring higher runtime than OpenAI Swarm due to richer coordination and memory overhead. On GSM8K, COSMIC attains$94\pm 1$% accuracy, exceeding the sparse multiagent debate (MAD) configuration reported in a Google Research publication. Empirically, performance peaks around five agents, reflecting a trade-off between diversity and coordination overhead.

We present a cloud-based approach for the design of interoperable electronic health record (EHR) systems. Cloud computing environments provide several benefits to all the stakeholders in the healthcare ecosystem (patients, providers, payers, etc.). Lack of data interoperability standards and solutions has been a major obstacle in the exchange of healthcare data between different stakeholders. We propose an EHR system - cloud health information systems technology architecture (CHISTAR) that achieves semantic interoperability through the use of a generic design methodology which uses a reference model that defines a general purpose set of data structures and an archetype model that defines the clinical data attributes. CHISTAR application components are designed using the cloud component model approach that comprises of loosely coupled components that communicate asynchronously. In this paper, we describe the high-level design of CHISTAR and the approaches for semantic interoperability, data integration, and security.

Applications over cellular networks now range from operator–consumer applications (e.g., mobile television, voice-over-ip, video conferencing), peer-to-peer applications (e.g., instant messaging), machine-to-machine applications (e.g., data telemetry and automotive applications), mobile web services (e.g., music and video streaming), and social networking applications. The current approach for developing mobile applications appears to focus on utilizing template-based application-development kits provided by platform developers (e.g., Google's Android, Apple's iOS, or Nokia's Symbian) to capture application designs and install them on the runtime platforms through use of code generators tied to particular versions of the platform. It is still unclear as to how an application developer (or network operator) conceptualizes the features of a mobile application in a platform-independent way, identifies its utility and explores its impact on the user, or further refines the choice of technology, platform, and mobility/interactivity requirements. This paper attempts to offer some guidelines, based on recent research in the industry and academia in these areas, toward the design and development of successful mobile applications that can utilize the capabilities of the next generation of cellular networks. We provide an overview of the growing trends of the rich multimedia and real-time mobile applications, including the diversity of application types, their impact on the enterprise and consumer, their traffic volumes, and their load and communication patterns. In addition to the overall trend analysis, we also study the design choices that are to be made, and how they are realized, and also describe how the platforms (client and server) may be implemented. Additionally, we focus on mobile video applications according to their communication characteristics and their distinct demands on the cellular network. We also present an analysis of device and network application programming interfaces (API) that form the basic building blocks for efficient and secure mobile application development of the future.

We present a novel framework, CloudView, for storage, processing and analysis of massive machine maintenance data, collected from a large number of sensors embedded in industrial machines, in a cloud computing environment. This paper describes the architecture, design, and implementation of CloudView, and how the proposed framework leverages the parallel computing capability of a computing cloud based on a large-scale distributed batch processing infrastructure that is built of commodity hardware. A case-based reasoning (CBR) approach is adopted for machine fault prediction, where the past cases of failure from a large number of machines are collected in a cloud. A case-base of past cases of failure is created using the global information obtained from a large number of machines. CloudView facilitates organization of sensor data and creation of case-base with global information. Case-base creation jobs are formulated using the MapReduce parallel data processing model. CloudView captures the failure cases across a large number of machines and shares the failure information with a number of local nodes in the form of case-base updates that occur in a time scale of every few hours. At local nodes, the real-time sensor data from a group of machines in the same facility/plant is continuously matched to the cases from the case-base for predicting the incipient faults-this local processing takes a much shorter time of a few seconds. The case-base is updated regularly (in the time scale of a few hours) on the cloud to include new cases of failure, and these case-base updates are pushed from CloudView to the local nodes. Experimental measurements show that fault predictions can be done in real-time (on a timescale of seconds) at the local nodes and massive machine data analysis for case-base creation and updating can be done on a timescale of minutes in the cloud. Our approach, in addition to being the first reported use of the cloud architecture for maintenance data storage, processing and analysis, also evaluates several possible cloud-based architectures that leverage the advantages of the parallel computing capabilities of the cloud to make local decisions with global information efficiently, while avoiding potential data bottlenecks that can occur in getting the maintenance data in and out of the cloud.

We propose a new family of high-rate space-time block codes called embedded orthogonal space-time (EOS) codes. The family is parameterized by the number of transmit antennas, which can be any positive integer, and by the rate, which can be as high as half the number of transmit antennas. The proposed codes are based on a new concept called embedding, whereby information symbols of a traditional space-time code are replaced by codewords from a second space-time code. The EOS codes use orthogonal designs as this second code, which induces orthogonality in an effective channel matrix and leads to reduced-complexity decoding. The EOS codes have lower decoding complexity than previously reported space-time codes for any number of transmit antennas, and for any rate. Furthermore, simulation results show that the EOS codes outperform previous constructions for certain number of antennas and certain rates, when performance is measured by error probability in quasistatic Rayleigh fading.

This paper introduces a new idea for implementing playback adaptation algorithms suitable for video streaming applications in wireless IP networks. The proposed algorithm depends on the accurate estimation of the expected latency of lost video packets. The latency is estimated according to the transport protocol being used, which in our case is TCP. The client uses this estimate in order to adjust the playback rate according to a new algorithm so that both buffer underflows and overflows can be avoided. The proposed algorithm in non-intrusive to the protocol stack since it does not require any modifications to the transport protocol. To validate the basic principles behind our idea, we provide a set of comparative experimental results with other streaming systems in terms of the rate of underflow events and the actual PSNR.

At the forefront of the recent advances in mobile networks is the development of sophisticated mobility management mechanisms that are usually based on Mobile IP and its derivatives. Based on these mobility management protocols, several studies that characterize transport protocol performance have been presented. In this work we move one step further, and present a joint performance evaluation model of TCP and TFRC, with the underlying IP-based mobility protocols. We develop stochastic models that can characterize the protocol performance during handoffs between heterogeneous wireless networks like WLAN, cellular, or WMAN. We present performance evaluation results for validating the developed models under a set of different handoff scenarios. The developed model can be utilized as a basis for further analytical evaluation of new mobility management protocols, allowing thus a fast and accurate comparison.

We present closed-form analytic expressions for the signal-to-noise ratio (SNR) of the downlink of 3G CDMA systems using root raised cosine (RRC) chip pulse shaping with arbitrary roll-off factors, quadrature phase shift keying (QPSK) modulation, and maximum ratio combining (MRC) RAKE receiver. The expressions obtained are also applicable to variable processing gain, chip rate, and data rate for orthogonal codes, random codes, and include adjacent channel interference, out-of-cell interference, and Gaussian noise. Moreover, the analysis applies to channel models with arbitrary path separation and delay spread and takes into account the self-interference. The mathematical derivation utilized is both insightful and straight-forward. The theoretical expressions produce results at a speed several orders of magnitude faster than Monte-Carlo simulations while closely matching them. Additionally, the so-called orthogonality factor, used for system-level simulations, is also derived in this work

We propose a novel end-to-end architecture for streaming H.264/AVC (advanced video coding) unicast video over the Internet. The proposed video streaming architecture is based primarily on a new transport layer protocol, the stream control transmission protocol (SCTP). We show that the network-friendly specification of H.264/AVC and the novel technical characteristics of SCTP, when coupled together are able to provide a highly adaptive and flexible system for unicast video streaming. With simulation results we prove (1) that our system is capable of maintaining good perceptual quality under various loss conditions and (2) that the proposed system maintains TCP-friendliness.

The advent of wireless ad-hoc networks has provided a cost effective way for exploiting mobile hosts in the absence of infrastructure. However, the full potential that these networks offer for performance enhancement still remains to be exploited. In this paper we exploit the existence of multiple paths between mobile hosts in an ad-hoc network. More specifically, we introduce a new scheme for load-balancing for mobile multihomed hosts that are part of an ad-hoc network. Our approach differs from previously proposed schemes since it consists an entirely end-to-end approach built on top of the stream control transmission protocol (SCTP), while the dynamic source routing (DSR) protocol is used for routing. We provide simulation results that prove the efficiency of our approach based on realistic mobility scenarios.

This paper describes the advantages of SyD (System on Mobile Devices), a middleware technology for mobile devices and e-services, in terms of technology and programming. Features of SyD are illustrated here through our prototype application, a complex communication system for a trucking fleet that operates an automated package delivery system. The fleet system has been implemented in three ways, with SOAP, with JDBC, and with SyD. Our implementation experience shows that SyD greatly simplifies coding by allowing heterogeneous devices, peer-to-peer communications, group transactions based on triggering events, and mobility support through proxies and directory service.

We present a number of modifications to the stream control transmission protocol (SCTP) recently adopted by IETF; they allow bandwidth aggregation over the multiple interfaces of a host. We show that it is possible to implement a number of algorithms for bandwidth aggregation, with only a small number of modifications to the basic SCTP. Our simulation results clearly depict the efficiency of our approach in terms of bandwidth utilization. Furthermore, we implement and evaluate a mechanism for identifying bottlenecks that are shared by flows from the same aggregate connection. Our purpose is to show that SCTP is a good candidate for building a practical protocol for bandwidth aggregation that is fair and supportive of TCP.

In this paper we evaluate the performance of the stream control transmission protocol (SCTP) in a wireless ad-hoc network environment under two routing protocols: DSR and AODV. Furthermore, we propose a cross-layer optimization to the SCTP and DSR protocols that handle pro-actively route failures. The novelty in the proposed mechanism consists in the identification of route failures which is done at the transport layer, allowing thus faster recovery. Our simulation results prove the efficiency of our approach for a variety of mobility scenarios.

In this paper, we present a mechanism for runtime updating of all kernel modules of a highly modular dynamic real-time operating system. Our approach can help solve the lack of adaptability, extensibility and flexibility of existing real-time operating systems. The dynamic real-time operating system will efficiently support a wide range of applications since any kernel module can be dynamically loaded at runtime to exactly suit the applications without necessitating a reboot of the system.

Two new fast algorithms for computing the inverse modified discrete cosine transform (IMDCT) as used in the adaptive spectral entropy coder (ASPEC) (Brandenburg et al. 1991) are proposed. A fixed-point error analysis is presented to determine the number of significant bits required for fixed-point implementations.

Most standards provide a generalized syntax and semantics framework for video coders, leaving the selection and optimization of the right parameter set (and lookup tables) to the implementation. The choice of the right parameter set that is suitable for a rich enough class of input sequences is, however, quite difficult. This difficulty is particularly amplified in the low-bit-rate video coding arena, where robust parameter sets are very important. We propose that robust parameter estimation, using the Taguchi (1979) methods, when applied to low-bit-rate video coding allows effective (near optimal) performance over a wide variety of input data streams. A number of experimental results confirm the improvement (via robustness) vis-a-vis conventional parameter estimation methods, and these methods promise a solution to the design of efficient parameter sets that support standards.

An oversampling technique that allows ternary (-1,0,1) filter coefficients is proposed for the implementation of two-band perfect reconstruction-quadrature mirror filter (PR-QMF) integer filterbanks. The optimization algorithm is based on the energy compaction. The resulting filterbank has two advantages: (i) the multiplier-free structure allows small VLSI area and low-power, and (ii) the filters use the whole given integer range leading to better objective performance compared to the previously published approaches.

We propose an open signal processing system design and implementation environment, BEEHIVE, that allows application developers to rapidly compose and debug functional specifications in a networked, distributed computing environment, and then later migrate the application (transparently) onto an embedded, distributed, computing hardware/software platform, with the capability to reconfigure (adaptively) the resources assigned to the application to meet the dynamic real-time requirements of the implementation. Recent developments in the area of virtual machines; broker-based, distributed, transportable computing; object-oriented programming methodologies, Java and its real-time extensions; reconfigurable and programmable hardware; approximate algorithms; adaptive-load and resource-management algorithms, are harnessed in this operating environment.

A method to automatically detect targets from sets of pixel-registered visual, thermal, and range images is outlined. It uses operations specifically designed to work on the different kinds of images to exploit the information given by each of them. Five features are used to distinguish the targets from the clutter: texture, brightness, temperature, surface planarity, and height. The results from individual detectors are then combined to improve the detection rate while reducing the number of false alarms. A morphological operation called "erosion of strength n" is also introduced and utilized as a powerful tool for removal of spurious information. The excellent results obtained for detection support the suitability of this approach for other ATR (automatic target recognition) problems.

A new adaptive lapped transform (ALT) structure is introduced to implement time-varying linear-phase paraunitary and biorthogonal filter banks. The new ALT is based on the generalized lapped transform (GLT). A time-varying filter bank is constructed through the factorization of the GLT into cascaded matrix stages. Variable-length LTs are subsequently generated by cascading a number of these matrix stages to build specific length filters. A fixed and variable block size embedded ALT image codec is presented.

A structure is presented for implementing perfectly reconstructing time-varying linear-phase paraunitary filterbanks (LPPUFB) using a novel adaptive lapped transform (ALT). The ALT is based on the generalized lapped orthogonal transform (GenLOT). A time-varying filterbank is constructed through the factorization of the GenLOT into cascaded matrix stages. Variable-length lapped transforms are subsequently generated by cascading a number of these matrix stages to build filters of specific length. An embedded ALT image codec is presented.

This paper proposes a novel design methodology and environment for digital signal processing application-specific integrated circuits (DSP-ASICs). The design methodology is based on systematic functional reuse. A library defines basic functional blocks that are automatically reused over multiple DSP applications. The defined design approach is utilized to rapidly prototype high performance DSP applications. A comparison with conventional ASIC and FPGA design, shows that the proposed approach bridges the gap between ASICs and FPGAs by offering high performance, low cost, and a short time-to-market.

As part of ARPA's RASSP program, Lockheed Martin ATL Laboratories has proposed a methodology for rapid prototyping using a plug and play approach. The methodology is based on encapsulation of the basic architectural elements in a design library. The encapsulation interface provides a uniform external view that is independent of the internal functionality of its component. In addition to encapsulating the functionality, the library also captures the low level I/O protocols along with relevant timing information both in synthesis and simulation views that are used for logic generation and the virtual prototyping phase of the development, respectively. The encapsulation is done via VHDL to provide a standard path to the existing synthesis tools. Given a library of encapsulated components, the net list for a functional hardware prototype can be obtained in a matter of hours. We examine this approach in detail in this paper and present some new results on efficient encapsulation.

All digital signal processing (DSP) circuits use analog front-end anti-aliasing filters to process sensor inputs. These analog filters have a number of disadvantages with respect to cost, power, stability, and ease of integration in VLSI. This paper proposes the use of all-digital oversampling front-end sensors (based on adaptive filters), resulting in a performance/cost-effective DSP application. Simulations for modem and speech applications illustrate the design tradeoffs.

The Rapid Prototyping of Application Specific Signal Processors (RASSP) US Department of Defense (ARPA/Tri-Services) initiative is intended to dramatically improve the way digital systems, particularly embedded digital signal processors, are designed, manufactured, upgraded, and supported. These DSP systems are complex, with typically one or more printed circuit boards, a variety of implementation technologies and interfaces, and a wide range of data rates. The target RASSP improvement is at least a fourfold (4/spl times/) reduction in the time to go from design concept to fielded system. Equivalent improvements in cost and quality are also targets. The motivation for the RASSP initiative is the pervasive need for affordable embedded signal processors throughout a wide range of DoD systems, signal processors that are state-of-the-art when they are fielded rather than when they are first defined. A number of programs are included in the RASSP initiative. This paper introduces the initiative from both technical and programmatic viewpoints.

The Rapid Prototyping of Application Specific Signal Processors (RASSP) program initiated by the Advanced Research Projects Agency (ARPA) has proposed a design process that is based on a new design methodology called Virtual Prototyping, wherein VHDL models of hardware components are integrated with application, control and diagnostic software to rapidly prototype complex embedded DSP systems. This paper discusses this new methodology, and compares the RASSP process with current design practice (circa 1993).

Digital signal processing (DSP) has made significant impact on the fields of telecommunications, medical technology, information storage and retrieval, military applications, remote sensing, and man-machine communications. The DSP Laboratory at Georgia Tech is heavily involved in research and education in this important area. This paper discusses the state-of-art DSP software and hardware design education (graduate, undergraduate, and continuing education) being offered by our laboratory, and outlines the guiding principles and rationale that continue to shape our closely intertwined research and education goals. Recent curriculum changes at Georgia Tech are also discussed.

The Department of Defense ARPA program for Rapid Prototyping of Application Specific Signal Processors (RASSP) exists to significantly improve the process by which embedded digital signal processors are developed (prototyped) and supported (maintained and upgraded). As used in the RASSP program, the term prototype signifies a system that is a precursor to a deployed system, but still meets all of the essential performance goals and is designed to facilitate maintainability and upgradability. Current practice in the design of embedded digital signal processors, as exemplified in the traditional waterfall design methodology, is examined and shortfalls in the design, methodology and supporting tools are identified. Opportunities for improving the traditional design practice are then identified and evaluated in terms of potential benefits, as well as impediments, to implementation, and adoption by the community.

Portable and mobile applications require low power consumption in addition to high throughput to prolong battery life in a VLSI implementation. The authors propose a novel configuration of a low-power portable communications (PCS) device that promises improved performance over very noisy band-limited channels through the use of a feedback channel and digital signal processing techniques that have their origins in sigma-delta modulation. Comparisons with PCM using raised-cosine pulses illustrate the design tradeoffs.>

The author reformulates the standard spotlight mode synthetic aperture radar (SAR) problem as a (/spl tau/,p) Radon transform. This results in a new algorithm for SAR that uses a linear chirp FM source signal that is a function of the geometry between the source and the object. While the close relation between computer-aided-tomography (CAT) and SAR has been understood for over a decade, no accompanying computational advantages could be realized. Recently, Kelley and Madisetti (1991), have proposed a new approach to image reconstruction from projections using the so-called fast Radon transform (FRT). He derives, in the (/spl tau/,p) domain, a FRT-based solution to the SAR problem. The FRT offers implicit interpolation and parallel processing advantages not found in conventional back projection operations. He proposes its efficient application in the computationally intensive problems in telecommunications and remote sensing, especially in military applications.>

An inversion scheme for reconstruction of images from projections based on the slope-intercept form of the discrete Radon transform is presented. A seminal algorithm for the forward and the inverse transforms proposed by G. Beylkin (1987) demonstrated poor dispersion characteristics for steep slopes and could not invert transforms based on nonlinear slope variations. By formulating the computation of a discrete computation of the continuous Radon transform formula, the authors explicitly derive fast generalized inversion methods that overcome the original shortcomings. The generalized forward (FRT) and inverse (IFRT) algorithms proposed are fast, eliminate interpolation calculations, and convert directly between a raster scan grid and a rectangular/polar grid in one step.

The use of realistic structural and behavioral level descriptions of digital signal multiprocessors to find rate optimal and processor optimal schedules for implementation on multiprocessor digital signal processor (DSP) systems is addressed. Unlike previous work, the digital signal multiprocessor compiler (DSMP-C1) explicitly considers the location of operands, the number of accumulators and registers, the size of on-chip and external memories, the size of communication buffers, and inter-processor communications. The algorithm implicitly retimes and pipelines the flow graph and generates overlapping PSSIMD schedules that are both rate and processor optimal. When compared to previously published scheduling algorithms, the DSMP-C1 algorithm produces significantly shorter iteration periods.>

An explicit relationship between the continuous and discrete time Radon transforms is derived. A generalized least-squares solution to the inversion problem is proposed, and a new inverse counterpart to the fast Radon transform (FRT) algorithm (IFRT) is derived. The authors' interest in the FRT algorithm stems from its application to the seismic inversion problem. A method of seismic migration based upon the wave equation solution in the Radon transform domain is derived using the FRT and IFRT. A number of concurrent VLSI architectures that find favor from the viewpoint of efficient implementation are described. The FRT is extended to three dimensions.>

The authors introduce the seismic imaging problem and formulate several algorithms suitable for concurrent VLSI implementation. They introduce an optimal VLSI architecture for 2-D and 3-D frequency-domain migration and propose a new method of matrix transposition for the 2-D parallel phase shift migration algorithm. They also propose a systolic array architecture for computation of a time-domain finite-difference migration algorithm.>

Seismic migration in the frequency domain using the phase-shift method is a computationally efficient solution of the wave equation. The authors introduce a parallel phase-shift algorithm. This parallel version is implemented on the NCUBE multiprocessing system to solve the wave equation in the frequency domain. Performance issues, some task partitioning strategies, and computation/communication tradeoffs are discussed.>

The authors study seismic migration algorithms in order to develop a general framework for analyzing such algorithms and to propose some coarse-grained paradigms for computation of parallel computers. They analyze a few typical examples of the various migration techniques existing in literature, and discuss methods to optimize their performance on distributed concurrently executing processors. The computational and communication requirements of the algorithms are discussed and diverse optimization techniques are proposed. The memory limitation, I/O bottleneck, and computational tradeoffs on hypercube multiprocessors are analyzed.>

In some adaptive filter applications, particularly echo cancellation, most of the tap coefficients in the convolution sum are essentially zero. We introduce and examine some No-Overkill Schemes (NOS) that eliminate the multiplications required for these zero coefficients by estimating which coefficients are nearly zero and setting them to exactly zero. In addition to reducing computation, this approach has the advantage of reducing the excess mean-square error, since the taps set to zero have no asymptotic fluctuation, Specifically, the NOS schemes work by dynamically partitioning the filter taps into active-taps and zero-taps. We estimate the time to convergence and the reduction in the excess mean-square-error due to the statistical fluctuation of the falter coefficients. Our results indicate that these No-Overkill Schemes significantly reduce the computational requirements with little or no increase in the time to convergence. Robustness of these schemes to a time varying channel is also considered. The NOS schemes increase the control overhead in the implementation of the adaptive filter, so additional examination of implementation tradeoffs for minimum cost is required.