Marshall Brinn

Genetic algorithms for complex, real-time scheduling

Real-time scheduling of large-scale problems in complex domains presents a number of difficulties for search and optimization techniques, including: large and complex search spaces; dynamically changing problems; and a variety of problem-dependent constraints and preferences. Genetic algorithms are well suited to such problems due to their adaptability and their effectiveness at searching large spaces. We have used genetic algorithms to solve real-world problems in areas such as field service scheduling, air crew scheduling and transportation scheduling. We discuss key aspects of our approach including: domain-specific chromosome representation and genetic operators; multi-objective evaluation function; heuristic initialization of the population; dynamic rescheduling; and cooperative interaction with human operators.

A Framework to Control Emergent Survivability of Multi Agent Systems

As the science of multi-agent systems matures, many developers are looking to deploy mission critical applications on distributed multi-agent systems (DMAS). Due to their distributed nature, designing survivable resource constrained DMAS is a serious challenge. Fortunately, the intrinsic flexibility of DMAS allows them to shift resources at runtime between dimensions of functionality such as security, robustness, and the primary application. In this paper we present an algebra for computing overall survivability from these dimensions of success, and a control infrastructure that leverages these degrees of freedom to make run-time adaptations at multiple hierarchical levels to maximize overall system survivability. We have implemented this survivability control infrastructure on the Cougaar agent architecture, and built a military logistics application that can survive in chaotic environments. Finally, we present results from assessing the performance of this application, and discuss the implications for future deployed DMAS.

Leveraging agent properties to assure survivability of distributed multi-agent systems

The nature of distributed multi-agent systems makes assuring their survivability under stress particularly challenging. However, the nature of distributed agent-based systems also brings the potential to address these particular challenges, and, indeed, to assure survivability to a degree beyond that possible in non-agent-based architectures. This extended abstract synopsizes a paper detailing approaches that are rooted in the essential properties of agent software architectures to assure the survivability of distributed agent-based systems. Specifically, the paper describes efforts under the DARPA UltraLog program to formulate a survivability argument based on properties of agent architectures. This extended abstract truncates many details from the original; interested readers are encouraged to contact the authors for the complete paper.

Extending the limits of DMAS survivability: the UltraLog project

Multiagent systems, being distributed and autonomous, provide the potential for more efficient, component-based development approaches. The massive parallelism intrinsic to these systems promises greatly enhanced performance. However, the price of these potential benefits is often paid in survivability. Distributed multiagent systems have typically failed to provide reliable, assured performance even in well-controlled environments. DARPA challenged the team to achieve this survivability in a way that would convince other potential customers that they could also make their DMAS applications survivable. In response, we developed and applied a generic methodology for survivable DMASs.

Reliable MAS performance prediction using queueing models

In this paper, we model a multi-agent system (MAS) in military logistics based on the systemic specifications of the capabilities and attributes of individual agents (TechSpecs). Assuring the survivability of the MAS that implements distributed planning and execution is a significant design-time and run-time challenge. Dynamic battlefield stresses in military logistics range from heavy computational loads (information warfare) to being destructive to infrastructure. In order to sustain and recover from damages to continuously deliver performance, a mechanism that distributes knowledge about the capabilities and strategies of the system is crucial. Using a queueing model to represent the network of distributed agents, strategies are developed for a prototype military logistics system. The TechSpecs contain the capabilities of the agents, play-books or rules, quantities to monitor, types of information flow (input/output), measures of performance (quality of service) and their computation methods, measurement points, defenses against stresses and configuration details (to reflect command and control structure as well as task flow). With these details, models could be dynamically developed and analyzed in real-time for fine-tuning the system. Using a Cougaar (DARPA agent framework) based model for initial parameter estimation and analysis, we obtain an analytical and a simulation model and extract generic results. Results indicate strong correlation between experimental and actual events in the agent society.

Enabling Campus Edge Computing Using GENI Racks and Mobile Resources

This paper presents the architecture of GENI edge cloud computing network in the form of compute and storage resources, a mobile 4G cellular edge and a high speed campus network connecting these components. This deployment is available across fifty campuses in the US, all interconnected via a nationwide Layer-2 network. We present these capabilities in the context of vehicular sensing and control applications running on police patrol cars on the Wayne State University campus allowing end-users and researchers to collect rich datasets for public safety surveillance, vehicle internal-state sensing and modeling, and emulating next generation connected vehicle technologies. In particular, the paper provides insights about the usefulness of local edge computing cloud infrastructure for novel connected vehicle applications with high sensitivity to latency and bandwidth.

Evolution of internal dynamics for neural network nodes

Most artificial neural networks have nodes that apply a simple static transfer function, such as a sigmoid or gaussian, to their accumulated inputs. This contrasts with biological neurons, whose transfer functions are dynamic and driven by a rich internal structure. Our artificial neural network approach, which we call state-enhanced neural networks, uses nodes with dynamic transfer functions based on n-dimensional real-valued internal state. This internal state provides the nodes with memory of past inputs and computations. The state update rules, which determine the internal dynamics of a node, are optimized by an evolutionary algorithm to fit a particular task and environment. We demonstrate the effectiveness of the approach in comparison to certain types of recurrent neural networks using a suite of partially observable Markov decision processes as test problems. These problems involve both sequence detection and simulated mice in mazes, and include four advanced benchmarks proposed by other researchers.

Trust as the Foundation of Resource Exchange in GENI

Researchers and educators in computer science and other domains are increasingly turning to distributed test beds that offer access to a variety of resources, including networking, computation, storage, sensing, and actuation. The provisioning of resources from their owners to interested experimenters requires establishing sufficient mutual trust between these parties. Building such trust directly between researchers and resource owners will not scale as the number of experimenters and resource owners grows. The NSF GENI (Global Environment for Network Innovation) project has focused on establishing scalable mechanisms for maintaining such trust based on common approaches for authentication, authorization and accountability. Such trust reflects the actual trust relationships and agreements among humans or real-world organizations. We describe here GENI’s approaches for federated trust based on mutually trusted authorities, and implemented via cryptographically signed credentials and shared policies.

Every agent a web server, every agent community an intranet…

The research community is struggling with the right approach to integrate emerging agent technology with the exploding web page- based Internet. This paper presents the approach used in Cognitive Agent Architecture (Cougaar) under DARPAs Advanced Logistics Project (ALP). In this effort, each agent was constructed with a build- in web server, supported by specialized API for easy creation of modular PlugIns which transforms agent internal object structures into XML and formatted HTML. In addition, the project has developed an efficient set of techniques for transparent proxies to allow any agent entry- point to serve as an Intranet portal to the whole community of agents. Further, as each agent holds only its part of the plan, the underlying infrastructure transparently supports traversal across the boundaries of one agents plan into that of a different agent. Thus the interfaces support a visualization of the global plan though physically it is a partitioned blackboard distributed among many agents on different machines. The benefit of this approach is that developers can easily build specialized user interfaces into the agent internal plans and state which is accessible through a standard browser by connecting to any agent URL in the society.

Genomic computing networks learn complex POMDPs

A genomic computing network is a variant of a neural network for which a genome encodes all aspects, both structural and functional, of the network. The genome is evolved by a genetic algorithm to fit particular tasks and environments. The genome has three portions: one for specifying links and their initial weights, a second for specifying how a node updates its internal state, and a third for specifying how a node updates the weights on its links. Preliminary experiments demonstrate that genomic computing networks can use node internal state to solve POMDPs more complex than those solved previously using neural networks.