Rinde R. S. van Lon

Evolutionary synthesis of multi-agent systems for dynamic dial-a-ride problems

In dynamic dial-a-ride problems a fleet of vehicles need to handle transportation requests within time. We research how to create a decentralized multi-agent system that can solve the dynamic dial-a-ride problem. Normally multi-agent systems are hand designed for each specific application. In this paper we research the applicability of genetic programming to automatically program a multi-agent system that solves dial-a-ride problems. We evaluated the evolved system by running a number of simulations and compared its performance to a selection hyper-heuristic. The results shows that genetic programming can be a viable alternative to hand constructing multi-agent systems.

RinSim: A Simulator for Collective Adaptive Systems in Transportation and Logistics

Engineering collective adaptive systems (CAS) is a challenging task. Concurrent systems, esp. when being large-scale, are known to be hard to design as the overall system behavior non-linearly results from local behavior and interactions. They are also hard to engineer and debug, as time dependent errors are often hard to reproduce. Simulation tools and environments are often used to assist in this task. From our experience in developing and using simulators for decentralized systems (in traffic, logistics and smart power grid management), we learned that a simulation environment should comply to the following quality criteria. First, from a software engineering point of view, a simulation environment itself must be designed up to the highest software quality standards - modularity, separation of concerns, test-driven development, guaranteed state consistency, etc. are particularly important quality criteria to ensure correctness, extensibility and manageability of the software. Second, the simulation environment must provide convenient support for using and extending the simulation environment, ease the visualization of solutions, and - since its use in scientific process - offer direct support for evaluating CAS through the set-up of experiments. In this paper, we present RinSim, an open source simulator that explicitly addresses these quality criteria, and targets the large family of transportation and logistics applications. RinSim separates the definition of the problem domain from the solution, has a modular design, is being developed in a test-driven way, etc. RinSim has been used and extended in a variety of cases within our research group, and served as the core platform in our educational program on multi-agent software development.

Delegate MAS for Large Scale and Dynamic PDP: A Case Study

Pickup and Delivery Problems (PDPs) have received significant research interest in the past decades. Their industrial relevance has stimulated the study of various types of solutions. Both centralized solutions, using discrete optimization techniques, as well as distributed, multi-agent system (MAS) solutions, have proven their merits. However, real PDP problems today are more and more characterized by (1) dynamism - in terms of tasks, service time, vehicle availability, infrastructure availability, and (2) their large scale - in terms of the geographical field of operation, the number of pickup and delivery tasks and vehicles. A combination of both characteristics brings unsolved challenges.

Optimizing agents with genetic programming : an evaluation of hyper-heuristics in dynamic real-time logistics

Dynamic pickup and delivery problems (PDPs) require online algorithms for managing a fleet of vehicles. Generally, vehicles can be managed either centrally or decentrally. A common way to coordinate agents decentrally is to use the contract-net protocol (CNET) that uses auctions to allocate tasks among agents. To participate in an auction, agents require a method that estimates the value of a task. Typically, this method involves an optimization algorithm, e.g. to calculate the cost to insert a customer. Recently, hyper-heuristics have been proposed for automated design of heuristics. Two properties of automatically designed heuristics are particularly promising: (1) a generated heuristic computes quickly, it is expected therefore that hyper-heuristics perform especially well for urgent problems, and (2) by using simulation-based evaluation, hyper-heuristics can create a ‘rule of thumb’ that anticipates situations in the future. In the present paper we empirically evaluate whether hyper-heuristics, more specifically genetic programming (GP), can be used to improve agents decentrally coordinated via CNET. We compare several GP settings and compare the resulting heuristic with existing centralized and decentralized algorithms based on the OptaPlanner optimization library. The tests are conducted in real-time on a dynamic PDP dataset with varying levels of dynamism, urgency, and scale. The results indicate that the evolved heuristic always outperforms the optimization algorithm in the decentralized multi-agent system (MAS) and often outperforms the centralized optimization algorithm. Our paper demonstrates that designing MASs using genetic programming is an effective way to obtain competitive performance compared to traditional operational research approaches. These results strengthen the relevance of decentralized agent based approaches in dynamic logistics.

Towards Systematic Evaluation of Multi-agent Systems in Large Scale and Dynamic Logistics

A common hypothesis in multi-agent systems (MAS) literature is that decentralized MAS are better at coping with dynamic and large scale problems compared to centralized algorithms. Existing work investigates this hypothesis in a limited way, often with no support for further evaluation, slowing down the advance of more general conclusions. Investigating this hypothesis more systematically is time consuming as it requires four main components: (1) formal metrics for the variables of interest, (2) a problem instance generator using these metrics, (3) (de)centralized algorithms and (4) a simulation platform that facilitates the execution of these algorithms. Present paper describes the construction of an instance generator based on previously established formal metrics and simulation platform with support for (de)centralized algorithms. Using our instance generator, a benchmark logistics dataset with varying levels of dynamism and scale is created and we demonstrate how it can be used for systematically evaluating MAS and centralized algorithms in our simulator. This benchmark dataset is essential for enabling the adoption of a more thorough and systematic evaluation methodology, allowing increased insight in the strengths and weaknesses of both the MAS paradigm and operational research methods.

Stress assessment of car-drivers using EEG-analysis

At TUDelft there is a project running on the assessment of the emotional state of car-drivers. Emotions are usual assessed by analysis of facial expressions and voice analysis. In this paper we report about assessment of emotions via EEG analysis. In a car simulation environment car-drivers, equipped with neurocaps, have to drive different tracks with and without billboards. Registration of billboards generates specific brain signals P300 and has an impact on the eye-blink rate. We report about the experimental design and results of analysis.

When do agents outperform centralized algorithms

Multi-agent systems (MAS) literature often assumes decentralized MAS to be especially suited for dynamic and large scale problems. In operational research, however, the prevailing paradigm is the use of centralized algorithms. Present paper empirically evaluates whether a multi-agent system can outperform a centralized algorithm in dynamic and large scale logistics problems. This evaluation is novel in three aspects: (1) to ensure fairness both implementations are subject to the same constraints with respect to hardware resources and software limitations, (2) the implementations are systematically evaluated with varying problem properties, and (3) all code is open source, facilitating reproduction and extension of the experiments. Existing work lacks a systematic evaluation of centralized versus decentralized paradigms due to the absence of a real-time logistics simulator with support for both paradigms and a dataset of problem instances with varying properties. We extended an existing logistics simulator to be able to perform real-time experiments and we use a recent dataset of dynamic pickup-and-delivery problem with time windows instances with varying levels of dynamism, urgency, and scale. The OptaPlanner constraint satisfaction solver is used in a centralized way to compute a global schedule and used as part of a decentralized MAS based on the dynamic contract-net protocol (DynCNET) algorithm. The experiments show that the DynCNET MAS finds solutions with a relatively lower operating cost when a problem has all following three properties: medium to high dynamism, high urgency, and medium to large scale. In these circumstances, the centralized algorithm finds solutions with an average cost of 112.3% of the solutions found by the MAS. However, averaged over all scenario types, the average cost of the centralized algorithm is 94.2%. The results indicate that the MAS performs best on very urgent problems that are medium to large scale.

Measures of dynamism and urgency in logistics

Dynamism was originally defined as the proportion of online versus offline orders in the literature on dynamic logistics. Such a definition however, loses meaning when considering purely dynamic problems where all customer requests arrive dynamically. Existing measures of dynamism are limited to either (1) measuring the proportion of online versus offline orders or (2) measuring urgency, a concept that is orthogonal to dynamism, instead. The present paper defines separate and independent formal definitions of dynamism and urgency applicable to purely dynamic problems. Using these formal definitions, instances of a dynamic logistic problem with varying levels of dynamism and urgency were constructed and several route scheduling algorithms were executed on these problem instances. Contrary to previous findings, the results indicate that dynamism is positively correlated with route quality; urgency, however, is negatively correlated with route quality. The paper contributes the theory that dynamism and urgency are two distinct concepts that deserve to be treated separately.

Evolved multi-agent systems and thorough evaluation are necessary for scalable logistics (position paper)

We consider logistics problems that are both dynamic and potentially large scale. A common way to create a scalable solution for logistics problems is to use multi-agent systems. In this paper we take two positions of a different nature: (1) evolutionary designed multi-agent systems are a promising approach to create scalable and performant solutions for logistics problems, (2) the multi-agent systems field does not prioritize evaluation enough, which hinders thorough scientific comparisons and prevents adoption in industry. We present arguments and refute common counterarguments for our position. Further, we discuss our present and upcoming efforts to realize our position.

Design of Evolvable Biologically Inspired Classifiers

Complex systems are emergent, self-organizing and adaptive systems. They are pervasive in nature and usually hard to analyze or understand. Often they appear intelligent and show favorable properties such as resilience and anticipation. In this paper we describe a classifier model inspired by complex systems theory. Our model is a generalization of neural networks, boolean networks and genetic programming trees called computational networks. Designing computational networks by hand is infeasible when dealing with complex data. For designing our classifiers we developed an evolutionary design algorithm. Four extensions of this algorithm are presented. Each extension is inspired by natural evolution and theories from the evolutionary computing literature. The experiments show that our model can be evolutionary designed to act as a classifier. We show that our evolved classifiers are competitive compared to the classifiers in the Weka classifier collection. These experiments lead to the conclusion that using our evolutionary algorithm to design computational networks is a promising approach for the creation of classifiers. The benefits of the evolutionary extensions are inconclusive, for some datasets there is a significant performance increase while for other datasets the increase is very minimal.