In the research project (Next-Generation Multi-Purpose Production Systems - Distributed Production Control) a distributed, actor-based system has been realised, that uses Lagrangian Relaxation for optimising Flexible Job Shop Scheduling with Transport Times (FJSSTT) problems. The design of the architecture builds on the actor model. This design allows to combine operations research with distributed computing and is driven by the mathematical formulation of the Lagrange Relaxation approach. Runtime experiments with the initial implementation of the architecture have been done. The performance of the multi actor-based implementation is compared to other approaches finding solutions to the 𝒩𝒫 -hard FJSSTT problem.
A team formation problem consists in finding an effective group of experts in a social network to accomplish a job with a minimum expenditure of energy and time. This problem has been transposed into the domain of multiagent systems to form a team of autonomous agents whose mission is to achieve a given goal. There is a wide range of such problems. This paper generalizes one of them by assigning explicit behaviors to agents whose tasks are equipped with multiple attributes. Their values are compared with preferences attached to the desired tasks of the goal. A synthesized controller realizes the goal by invoking tasks of a subset of the available agents, called a composition in this paper. Furthermore, utility values are assigned to compositions and robustness is considered to be an important property of a team to prevent its deterioration when one or more of its agents fail. Finding a robust team that satisfies the goal's preferences with better utility values for compositions constitutes a difficult optimization problem. The proposed method to solve this problem consists in three phases: controller synthesis with filtering on tasks with respect to some qualitative preferences, composition ranking based on their fitness, and multiobjective mathematical optimization.
This paper describes and compares two of the algorithms for indoor localization that are implemented in the localization add-on module for JADE. Described algorithms perform localization of agents running on smart devices in known indoor environments using only received WiFi signals from access points. First, distance estimates from access points are computed using received signal strength in routinary network discovery. Then, computed distance estimates are used to generate estimates of the position of the smart device that hosts the agent using one of described algorithms. The first algorithm, known as two-stage maximum-likelihood algorithm, is a well-known technique and it is considered a point of reference to evaluate the performance of other algorithms. The second algorithm, which has been recently introduced to overcome numerical-instability problems of classic geometric algorithms, works by turning localization into an optimization problem which is effectively solved using particle swarm optimization. In order to show the applicability of the proposed algorithms, the last part of the paper shows experimental results obtained in an illustrative indoor scenario, which is representative of envisioned applications.
Ever increasing competition is driving the efforts to improve productivity throughout nearly all domains. In the manufacturing context, digitalization of value networks and creation of autonomous, self-optimizing systems – a vision coined 'Industrie 4.0 – is an approach that promises competitive edge over other players. One field in which this vision could lead to great productivity potentials is order scheduling and sequencing in high variety, high volume manufacturing businesses like the automobile industry. A viable technology to realize the expected gains in productivity are software agents and multi-agent systems, since they provide autonomy, flexibility, adaptiveness, and robustness to unforeseeable events. This paper proposes an agent-based control architecture that enables communication between resources and customer orders within a car body shop, so that they can negotiate the best alternative schedule and order sequence in case of disturbances. The proposed architecture allows improvement of overall production system performance in terms of output, resource utilization, delivery reliability and others. Further, the paper describes the implementation and simulation of the multi-agent system with JADE framework and discusses the simulation results, which show that significant productivity leaps can be achieved.
Ontology is one of the popular models for knowledge representation, and Web Ontology Language (OWL) is the current industry standard for supporting ontology in semantic web and knowledge encoding for various application domains including healthcare and cyber-security. But ontology basically only supports one relation “is-a” between the classes. Even though OWL introduced object properties to emulate other relations, it lacks effective support for fundamental relations like part-of, which is very popular in engineering knowledge, and the temporal relation, which underpins all algorithms for most computer science knowledge. This paper introduces our minimal syntax extension to OWL to allow domain experts to declare and apply custom relations with various mathematical properties, and our extension to Stanford University’s Protégé project so that it can be used to encode intuitively knowledge with custom relations, and our PaceJena project that can use the extended OWL documents to empower knowledge-driven decision making in software agents. Important use cases illustrate how this approach supported effective drug side-effect detection, efficient software diagnostic message pattern specification and detection, and an intelligent online tutoring system that supports effective cyberlearning with knowledge navigation, specialization and generalization as well as assessment-based learning.
This paper studies compromise, which is the tendency of agents to move their opinions towards those of agents they interact with, trying to reach consensus. Compromise is one of the most important phenomena in the study of opinion dynamics, and this paper presents two analytic models to study it. First, agents are considered deterministic and a preliminary model of the effects of compromise is derived. Then, the model is generalized to give agents some level of autonomy by modelling their behaviour in terms of a stochastic process. Both models are analytic and they can be used to study collective properties of multi-agent systems starting from the details of single interactions among agents. Analytic results concerning the conservation of the average opinion for both models are verified by simulation in the last part of the paper.
Medical diagnosis has always been a crucial and sophisticated matter, and despite its remarkable progresses, a reliable, cost-efficient, and fast computer-based medical diagnosis is still a challenge. There are two main types of computerized medical diagnosis systems: knowledge-based and non-knowledge-based systems. While the challenge of scalability and maintainability are the main shortcomings of the first group, the fact that the non-knowledge-based systems cannot explain the reasons for their conclusions makes them less appealing too. Moreover, even the most advanced systems fail to help the user in providing the right input. This work discusses the feasibility of the use of Holonic Multi-Agent Systems (HMASs) to tackle this problem, by performing differential diagnosis (DDx), that can improve diagnostic accuracy, and moreover guide the user in providing a more comprehensive input. The Holonic Medical Diagnosis System (HMDS), as a Multi-Agent System (MAS), offers the necessary reliability and scalability. By using Machine Learning (ML) techniques, it can also be self-adaptable to new findings. Furthermore, since it aims to perform DDx and tends to present the most likely diagnoses, the reasoning behind its output is also always implicitly recognizable. While theHMAS approach to DDx is the practical contribution of this work, the introduction of the ML techniques that support its functionality and dynamics is its theoretical contribution. Swarm Q-learning, as an off-policy reinforcement learning, is shown to be a perfect solution to this problem, and the Holonic-Q-learning technique is proposed, which can in general also be applied to any HMAS.
The Standard Vicsek Model and a popular variant—using topological neighbour interactions—are widely used models for studying flocking phenomena in the natural world. It is capable of demonstrating the ordered and disordered states of real world flocks by tuning a temperature variable η , where high η corresponds to the disordered state. Here we show that the ordered state attained at low η is not stable over indefinite time periods raising implications for simulations and settling times. Additionally, we show that the loss of coherency in the metric case is reversible, while it is permanent in topological case.
We present a method for learning to interpret and understand foreign agent communication. Our approach is based on casting the contents of intercepted opponent agent communication to a bit-level representation and on training and employing deep convolutional neural networks for decoding the meaning of received messages. We empirically evaluate our method on real-world data acquired from the multi-agent domain of robotic soccer simulation, demonstrating the effectiveness and robustness of the learned decoding models.
Social media like Facebook, Twitter, or Google+ have become important communication channels. Nonetheless, the distribution and dynamics of that communication make it difficult to analyze and understand. To overcome this, we propose an agent architecture for modeling and simulating user behavior to analyze communication dynamics in social media. Our agent decision-making method utilizes sociological actor types to represent motivations of media users and their impact on communicative behavior. We apply this concept to a simulation of real world Twitter communication accompanying a German television program. Our evaluation shows that the agent architecture is capable of simulating communication dynamics in human media usage.
Paradigms in modern production are shifting and pose new demands for optimization techniques. The emergence of new, versatile, reconfigurable and networked machines enables flexible manufacturing scenarios which require, in particular, planning and scheduling methods for cyber-physical production systems to be flexible, reasonably fast, and anytime. This paper presents an approach to flexible job-shop manufacturing scheduling with agent-based simulated trading, called shopST. Aspects of real manufacturing scheduling problems form the basis for a physical decomposition of the planning system into agents. The initial schedule created by the agents in shopST through reactive negotiation is successively improved through the exchange of resource binding constraints with an additional market agent. shopST is evaluated in comparison to selected other different solution approaches to flexible job-shop scheduling.
Inspired by the theory of practical reasoning, Belief-DesireIntention (BDI) agents are perhaps the most well-known type and architecture of cognitive agents. Such agents can reason about their environment and perform complex plans to bring about their objectives and goals. Within the context of ever-changing environments though, one desirable feature for agents is that of learning, implemented in BDI agents as Intentional Learning, a framework focused on the monitoring of the mental states to include learning as part of the agent goals. In this paper, we consider and develop intentional learning within the Jason BDI framework for agents focused on a plan acquisition strategy addressing the cases of learning plans composed of one action, sequences or a repetition of actions that allow an agent to improve its behaviour at run-time. This is done at the pure BDI agent level, the repertoire of plans is directly updated without using external planning tools. We take as a testbed the simple vacuum cleaning environment and how new plans are acquired for accomplishing tasks of different level of complexity: escape from tunnel-like paths and wall-following. Furthermore, we integrate in a novel way the use of NetLogo as an environment to locate Jason agents, maintaining a clear delineation between decision making and action in the environment with the decision-making firmly anchored within the BDI agent's reasoning cycle.
Multi-goal pathfinding (MGPF) is a problem of searching for a path between an origin and a destination, which allows a set of goals to be satisfied. We are interested in MGPF in ubiquitous environments that are composed of cyber, physical and social (CPS) entities from connected objects, to sensors and to people. Our approach aims at exploiting data from various resources such as CPS entities and the Web to solve MGPF. However, accessing resources creates overheads – specifically latency affecting the efficiency of the approach. In this paper, we present a collaborative multi-agent search model that addresses the latency problem. The model handles the process of accessing resources such that agents are not blocked while data from resources are being processed and transferred. Agents search concurrently and collaboratively on different parts of the search space. The model exploits the knowledge and structure of the search space to distribute the work among agents and to create an agent network facilitating agent communications as well as separating the search from the communications. To evaluate our model, we apply it in uniform cost search, creating a collaborative uniform cost algorithm. We compare it to the original algorithm. Experiments are conducted on search spaces of various sizes and structures. In most cases, collaborative uniform cost is shown to run significantly faster and scale better in function of latency as well as graph size.
Current energy grids are moving toward utilization of renewable and non-polluting energy sources. Micro-grids, as an emerging means for a localized management, supervision, and control of energy production and consumption are changing the traditional centralized grid topology, making it more distributed and autonomous. However, the fluctuating nature of renewable energy systems make the energy demand control very complex. Hence, one of the challenges in Micro-grid energy control and management is to handle any deviation from the prior forecasted power generation/consumption by optimizing the usage of storage and backup generation units in a way that preserves the users' convenience level. The majority of the proposed optimization approaches only use the centralized load shedding schemes, neglecting the effect of inconvenience it may cause to the users. In this paper, we propose a Multi-agent based decentralized algorithm for a residential grid-connected Microgrid. The focus of our work is on how to handle possible power imbalance situations with the help of an Autonomous Decentralized Multi-agent approach consisting of user agents, storage agent, and grid agent considering the users' consumption preferences as an important factor in the decision making. We investigate the application of our proposed algorithm over a PV-based Microgrid scenario.
The Braess' paradox is a well-known problem associated with route choice and traffic distribution. Agent-based simulations that investigate this paradox typically model driver's behaviour using reactive agent architectures, which simplify and abstract an inherently complex behaviour. The BDI architecture is an alternative widely used in multiagent systems, which has not been evaluated as a suitable solution to deal with this problem. We thus in this paper detail an empirical evaluation of the BDI architecture, enhanced with a learning-based plan selection, to address the Braess' paradox. We describe the results of two simulations configured to reproduce the paradox behaviour. Results indicate that agents are able to soften the effects of the Braess' paradox using only local information, as opposed to existing alternatives, including when the environment is dynamic.
In this work we propose and investigate the use of collaborative reinforcement learning methods for resolving demand-capacity imbalances during pre-tactical Air Traffic Management. By so doing, we also initiate the study of data-driven techniques for predicting multiple correlated aircraft trajectories; and, as such, respond to a need identified in contemporary research and practice in air-traffic management. Our simulations, designed based on real-world data, confirm the effectiveness of our methods in resolving the demand-capacity problem, even in extremely hard scenarios.
This work approaches the question whether or not agents are able to learn the personality of a human during interaction. We develop two agent-models to learn about the personality of humans during repeatedly played rounds in the Colored Trails Game. Human personality is described using a psychological theory of personality traits known as the Five-Factor Model. The results show that some characteristics of a personality can be learned more accurately than others. The work extends the state-of-the-art in that it does not follow a supervised learning approach requiring existing data sets.
Personality is one of the central elements determining the behaviour of humans. It influences other cognitive mechanisms such as emotions and moods and thus effects attention and actions. However, in the literature about cognitive agents, work that investigates the effects of personality is rare and somewhat disconnected. Bridging this gap represents one step towards conceptualising human behaviour in software agents, e.g. for resource-bounded agents in highly-dynamic environments or for virtual humans with realistic behaviour. The integration of personality in agents also requires its integration into reasoning processes used in agent-based systems. In this paper, we propose a formalisation that enables reasoning about the effectsand state of personality. This formalisation is integrated into the ‘ ℒ ogic 𝒪 f ℛ ational 𝒜 gents’ ( ℒ𝒪ℛ𝒜 ) and is the foundation for reasoning about the personality of other agents and the influence of personality on the action selection process.
Due to new energy transition policies fossil- and partially also nuclear based power production has been replaced by usually much more volatile renewable energy production. Volatility here means that it is substantially more difficult to keep energy production and consumption in balance. Such a challenge can be tackled by seamlessly integrating modern power storage technologies, such as batteries. This paper focuses on the role of battery storage providers to reveal their profitability and balancing potentials from the perspective of all parties, involved in the balancing process. Batteries are managed by broker agents, which are analyzed in the PowerTAC marketplace simulation environment. We first describe the Vickrey–Clarke–Groves auction mechanism to demonstrate how its pricing mechanism provides incentives for participants. Afterwards, we analyze the trading behaviors of different balancing settings to benchmark profitability levels of battery storage providers. We employ a broker agent for each setting so that variants publish a battery storage tariff with different up-regulation and down-regulation prices. The results show that battery storage providers provide extra profit for brokers if exploited strategically in the balancing market. Additionally, they help stabilizing the grid with up and down regulations.
One of the most important challenges of this decade is the Internet of Things (IoT) that pursues the integration of real-world objects in the virtual world of the Internet. One property that characterises IoT systems is that they have to react to variable and continuous changes. This means that IoT systems need to work as self-managed systems to effectively manage context changes. The autonomy property inherent to software agents makes them a suitable choice for developing self-managed IoT systems. By embedding agents in the devices that compose the IoT is possible to realize a decentralized system with self-management capacities. However, in this scenario new problems arise. Firstly, current agent development approaches lack mechanisms to deal with the heterogeneity present in the IoT domain. Secondly, agents must simultaneously deal with potentially conflicting changes in their behaviour, concerning self-management and application goals. In order to afford these challenges we propose to use an approach based on Dynamic Software Product Lines (D-SPL) and preference-based reasoning. The D-SPL provides to the preference-based reasoning of the agent with the necessary information to adapt its behaviour at runtime making a trade-off between the self-management of the system and the accomplishment of its application goals.
