Kyriakos C. Chatzidimitriou

A robust agent design for dynamic SCM environments

The leap from decision support to autonomous systems has often raised a number of issues, namely system safety, soundness and security. Depending on the field of application, these issues can either be easily overcome or even hinder progress. In the case of Supply Chain Management (SCM), where system performance implies loss or profit, these issues are of high importance. SCM environments are often dynamic markets providing incomplete information, therefore demanding intelligent solutions which can adhere to environment rules, perceive variations, and act in order to achieve maximum revenue. Advancing on the way such autonomous solutions deal with the SCM process, we have built a robust, highly-adaptable and easily-configurable mechanism for efficiently dealing with all SCM facets, from material procurement and inventory management to goods production and shipment. Our agent has been crash-tested in one of the most challenging SCM environments, the trading agent competition SCM game and has proven capable of providing advanced SCM solutions on behalf of its owner. This paper introduces Mertacor and its main architectural primitives, provides an overview of the TAC SCM environment, and discusses Mertacors performance.

Agent Mertacor: A robust design for dealing with uncertainty and variation in SCM environments

Supply Chain Management (SCM) has recently entered a new era, where the old-fashioned static, long-term relationships between involved actors are being replaced by new, dynamic negotiating schemas, established over virtual organizations and trading marketplaces. SCM environments now operate under strict policies that all interested parties (suppliers, manufacturers, customers) have to abide by, in order to participate. And, though such dynamic markets provide greater profit potential, they also conceal greater risks, since competition is tougher and request and demand may vary significantly in the quest for maximum benefit. The need for efficient SCM actors is thus implied, actors that may handle the deluge of (either complete or incomplete) information generated, perceive variations and exploit the full potential of the environments they inhabit. In this context, we introduce Mertacor, an agent that employs robust mechanisms for dealing with all SCM facets and for trading within dynamic and competitive SCM environments. Its efficiency has been extensively tested in one of the most challenging SCM environments, the Trading Agent Competition (TAC) SCM game. This paper provides an extensive analysis of Mertacor and its main architectural primitives, provides an overview of the TAC SCM environment, and thoroughly discusses Mertacors performance.

Data mining for agent reasoning: A synergy for training intelligent agents

The task-oriented nature of data mining (DM) has already been dealt successfully with the employment of intelligent agent systems that distribute tasks, collaborate and synchronize in order to reach their ultimate goal, the extraction of knowledge. A number of sophisticated multi-agent systems (MAS) that perform DM have been developed, proving that agent technology can indeed be used in order to solve DM problems. Looking into the opposite direction though, knowledge extracted through DM has not yet been exploited on MASs. The inductive nature of DM imposes logic limitations and hinders the application of the extracted knowledge on such kind of deductive systems. This problem can be overcome, however, when certain conditions are satisfied a priori. In this paper, we present an approach that takes the relevant limitations and considerations into account and provides a gateway on the way DM techniques can be employed in order to augment agent intelligence. This work demonstrates how the extracted knowledge can be used for the formulation initially, and the improvement, in the long run, of agent reasoning.

Information agents cooperating with heterogenous data sources for customer-order management

As multi-agent systems and information agents obtain an increasing acceptance by application developers, existing legacy Enterprise Resource Planning (ERP) systems still provide the main source of data used in customer, supplier and inventory resource management. In this paper we present a multi-agent system, comprised of information agents, which cooperates with a legacy ERP in order to carry out orders posted by customers in an enterprise environment. Our system is enriched by the capability of producing recommendations to the interested customer through agent cooperation. At first, we address the problem of information workload in an enterprise environment and explore the opportunity of a plausible solution. Secondly we present the architecture of our system and the types of agents involved in it. Finally, we show how it manipulates retrieved information for efficient and facile customer-order management and illustrate results derived from real-data.

Adaptive reservoir computing through evolution and learning

The development of real-world, fully autonomous agents would require mechanisms that would offer generalization capabilities from experience, suitable for a large range of machine learning tasks, like those from the areas of supervised and reinforcement learning. Such capacities could be offered by parametric function approximators that could either model the environment or the agents policy. To promote autonomy, these structures should be adapted to the problem at hand with no or little human expert input. Towards this goal, we propose an adaptive function approximator method for developing appropriate neural networks in the form of reservoir computing systems through evolution and learning. Our neuro-evolution of augmenting reservoirs approach comprises of several ideas, successful on their own, in an effort to develop an algorithm that could handle a large range of problems, more efficiently. In particular, we use the neuro-evolution of augmented topologies algorithm as a meta-search method for the adaptation of echo state networks for handling problems to be encountered by autonomous entities. We test our approach on several test-beds from the realms of time series prediction and reinforcement learning. We compare our methodology against similar state-of-the-art algorithms with promising results.

Transferring evolved reservoir features in reinforcement learning tasks

The major goal of transfer learning is to transfer knowledge acquired on a source task in order to facilitate learning on another, different, but usually related, target task. In this paper, we are using neuroevolution to evolve echo state networks on the source task and transfer the best performing reservoirs to be used as initial population on the target task. The idea is that any non-linear, temporal features, represented by the neurons of the reservoir and evolved on the source task, along with reservoir properties, will be a good starting point for a stochastic search on the target task. In a step towards full autonomy and by taking advantage of the random and fully connected nature of echo state networks, we examine a transfer method that renders any inter-task mappings of states and actions unnecessary. We tested our approach and that of inter-task mappings in two RL testbeds: the mountain car and the server job scheduling domains. Under various setups the results we obtained in both cases are promising.

Enhancing agent intelligence through evolving reservoir networks for predictions in power stock markets

In recent years, Time Series Prediction and clustering have been employed in hyperactive and evolving environments ---where temporal data play an important role--- as a result of the need for reliable methods to estimate and predict the pattern or behavior of events and systems. Power Stock Markets are such highly dynamic and competitive auction environments, additionally perplexed by constrained power laws in the various stages, from production to transmission and consumption. As with all real-time auctioning environments, the limited time available for decision making provides an ideal testbed for autonomous agents to develop bidding strategies that exploit time series prediction. Within the context of this paper, we present Cassandra , a dynamic platform that fosters the development of Data-Mining enhanced Multi-agent systems. Special attention was given on the efficiency and reusability of Cassandra , which provides Plug-n-Play capabilities, so that users may adapt their solution to the problem at hand. Cassandra s functionality is demonstrated through a pilot case, where autonomously adaptive Recurrent Neural Networks in the form of Echo State Networks are encapsulated into Cassandra agents, in order to generate power load and settlement price prediction models in typical Day-ahead Power Markets . The system has been tested in a real-world scenario, that of the Greek Energy Stock Market.

From requirements to source code: a Model-Driven Engineering approach for RESTful web services

During the last few years, the REST architectural style has drastically changed the way web services are developed. Due to its transparent resource-oriented model, the RESTful paradigm has been incorporated into several development frameworks that allow rapid development and aspire to automate parts of the development process. However, most of the frameworks lack automation of essential web service functionality, such as authentication or database searching, while the end product is usually not fully compliant to REST. Furthermore, most frameworks rely heavily on domain specific modeling and require developers to be familiar with the employed modeling technologies. In this paper, we present a Model-Driven Engineering (MDE) engine that supports fast design and implementation of web services with advanced functionality. Our engine provides a front-end interface that allows developers to design their envisioned system through software requirements in multimodal formats. Input in the form of textual requirements and graphical storyboards is analyzed using natural language processing techniques and semantics, to semi-automatically construct the input model for the MDE engine. The engine subsequently applies model-to-model transformations to produce a RESTful, ready-to-deploy web service. The procedure is traceable, ensuring that changes in software requirements propagate to the underlying software artefacts and models. Upon assessing our methodology through a case study and measuring the effort reduction of using our tools, we conclude that our system can be effective for the fast design and implementation of web services, while it allows easy wrapping of services that have been engineered with traditional methods to the MDE realm.

Redefining the Market Power of Small-Scale Electricity Consumers through Consumer Social Networks

Energy markets have undergone important changes at the conceptual level over the last years. Decentralized supply, small-scale production and smart grid optimization and control are the new building blocks. These changes offer substantial opportunities for all energy market stakeholders, some of which however, remain largely unexploited. Small-scale consumers, as a whole, account for significant amount of energy in current markets (up to 40%), as individuals though their consumption is trivial, and their market power practically non-existent. Thus, it is necessary to assist small-scale energy market stakeholders combine their market power. Within the context of this work we propose Consumer Social Networks (CSNs) as a means for achieve the objective. We present a simulation environment for the creation of CSNs and provide a proof of concept on how CSNs can be formulated based on various criteria. Each cluster in a CSN may be treated as a nontrivial stakeholder with specific characteristics that can actively affect energy market pricing policies. We also show provide an indication on how demand response programs designed based on targeted incentives may lead to energy peak reductions.

A Zeroth-Level Classifier System for Real Time Strategy Games

Real Time Strategy games (RTS) provide an interesting test bed for agents that use Reinforcement Learning (RL) algorithms. From an agents point of view, RTS games constitute a Markovian, partially observable and dynamic environment with a huge state space. In this paper, we present an agent that uses a Zeroth-level Classifier System (ZCS) in order to construct winning policies for this type of games. We also combine ZCS with the replacing traces method in an attempt to improve the behaviour of our agent. We tested the learning abilities of our agent against a static opponent. For the evaluation of our agent, we compare its results with those of a random-acting agent and an agent that uses the SARSA RL algorithm. Results are encouraging since, our ZCS agent managed to outperform the SARSA agent. On the other hand, applying replacing traces to ZCS did not yield the expected results.