inicio blog investigación publicaciones proyectos cv github
EN

PhD Thesis

Abstract

“Improving the Decision Support in Shop Floor Operations by Using Agent-based Systems and Visibility Frameworks” Pablo García Ansola — Universidad de Castilla-La Mancha (UCLM), 2012

Multi-agent system (MAS) models suffer from an excessive dependence on up-to-date field information and on complex interaction protocols. This dissertation proposes MAS-DUO, a theoretical and experimental agent-based decision support architecture designed to align shop floor operations with strategic policies while incorporating feedback from Radio Frequency Identification (RFID) visibility frameworks (EPCglobal / EPCIS). The architecture was tested in the ground handling operations of the Ciudad Real Central Airport and in a simulated logistics centre at Autolog Labs (UCLM).

The thesis was awarded the International Doctor distinction (cum laude) by the Universidad de Castilla-La Mancha (UCLM) in June 2012. A research stay at the Institute for Manufacturing (IfM), University of Cambridge (Aero-ID project), provided the international dimension required for that designation.

Research Problem and Motivation

Mass customisation, JIT/agile manufacturing and the airport Common Use model make shop floor control highly dynamic, while traditional hierarchical and CIM-style control remains rigid. The thesis identifies three information gaps at the shop floor:

  1. Knowing when a resource is actually free or busy.
  2. Estimating the duration of tasks whose timing depends on the context (e.g. in-block time estimates for incoming flights are highly volatile).
  3. Coordinating heterogeneous actors and propagating the strategic policy down to every operational decision.

Objectives

  • Main objective: accurate shop floor decisions through an intelligent-product architecture based on MAS that processes field information and aligns it with strategic policies.
  • Sub-objectives: (1) integrate field information into decision making through visibility frameworks; (2) align decisions through interaction protocols and information processing.
  • Technical objectives: a real-time image of the shop floor, dynamic negotiation protocols that preserve the independence of each party, service discovery, and integration with ERP/CRM/WMS systems as well as PLCs and robots.

Methodology

The agent system was specified by combining CommonKADS, GAIA and AUML (organisation, agent, task, learning and coordination models). Coordination uses variations of the FIPA Contract Net protocol. The prototype was implemented in Java on JADE (FIPA-ACL messaging) with a custom BDI module built on the JADE agent core, Fosstrak as EPCIS repository, MySQL, and a web-service interface to PLC ladder functions.

Technical Architecture: MAS-DUO

Following interviews with companies that rejected a single integrated system, MAS-DUO splits the agent space into two horizontal platforms of equal rank that only exchange results, policy parameters and instructions:

  • Physical BDI agent platform — close to the hardware (PLC, RFID). It contains resource agents (stairs, belts, tractors, forklifts…) and product agents (a flight, a pallet). Beliefs are updated through EPCIS subscriptions, so each agent reasons over RFID events.
  • IS BDI agent platform — interfaces with enterprise systems through three agents: an ERP agent, a CRM agent and an expert system agent. They validate the proposals coming from the physical platform and, if they reject them, send new global policy parameters.

State model. The state of a product is defined by the Where? (read point) and Why? (business step) of EPC events: States(p) = ReadPoint(p) × BusinessStep(p) (Eq. 4), and a task is a sequence of such states (Eq. 5).

Markov decision process. Each transition is evaluated with a weighted reward over operational criteria (Eq. 6 and 8):

$$R(s, s’) = A \cdot \text{Time}(s, s’) + B \cdot \text{Cost}(s, s’) + C \cdot \text{QoS}(s, s’) + D \cdot \text{Energy}(s, s’)$$

The weights (A, B, C, D) are the global policy parameters (Eq. 9), set by the IS platform. Resources answer a product’s request with their best reward, the product chooses, and the IS platform validates the choice. The thesis distinguishes a static policy (no RFID feedback; useful for simulation, training and resource sizing) from a dynamic policy fed by EPC events, and detects lack of visibility when a transition takes much longer than its average (Eq. 10).

Key Contributions

  1. Split organisation of the agent space into physical BDI agents (the intelligent product) and IS BDI agents, keeping physical control and information systems independent.
  2. Integration of an MDP into shop floor decisions driven by EPC/RFID states, without a supervisor agent.
  3. Policy alignment mechanism: global policy parameters negotiated with ERP, CRM and expert-system agents and propagated to every operational decision.
  4. Hardware integration lessons: a flexible agent–PLC interface (Siemens S7 network, PROFIBUS-DP) and RFID integration, together with the practical observation that deployment requires programmers skilled in both MAS and PLCs.

Case Studies and Results

Ciudad Real Central Airport — ramp ground handling (Chapter 4.1)

  • ZigID visibility infrastructure: a ZigBee (IEEE 802.15.4) mesh wireless sensor network with RFID reader nodes covering the parking stands.
  • Product agent = a flight (Iberia B737-800). Flights have 10 business steps (the 8 IATA ground handling functions plus airport-specific steps) × 4 read points = 40 states. Resources have 5 read points × 4 business steps = 20 states (Eq. 11).
  • Reward function fixed by the airport management staff (Eq. 12): R(s,s') = 0.5 Time(s,s') + 0.4 Cost(s,s') + 0.1 Energy(s,s'). QoS was excluded so that no airline would be given preference.
  • Evaluation: a demonstrator reproducing a standard operating day, comparing the number of material ground handling resources required by MAS-DUO against a FIFO allocation with indivisible handling teams (1–6 flights/hour; detailed analysis at 4 flights/hour).
  • Results (Table 4): an average reduction of 15.89 % in material resources in the just on time configuration, and 31.25 % when the policy accepts five minutes of delay.
  • Scope: these figures correspond to the static policy in a non-stochastic environment. The thesis leaves the quantitative impact of the dynamic, RFID-driven policy and of human resources for future work.

Autolog Labs logistics centre (Chapter 4.2)

A physical 1:20 scale mock-up (docks, conveyors, RFID readers, a Motoman HP3 robot and a 36-position rack) controlled by a network of seven Siemens PLCs, plus a 3D simulation of the distribution centre. This case study is qualitative: it analyses hardware integration, development problems and deployment constraints rather than productivity metrics.

Limitations and Future Work

The thesis itself highlights the need for MAS/PLC specialists, vendor-dependent PLC integration, the effort of filling in reward matrices manually, and an ROI that is negative over three years when only material resources are considered. Future lines include online outsourcing between companies, standard interfaces (IEC 61499, Internet of Things), POMDPs, self-diagnosis and deadlock management, a web adaptation (REST/JSON, Web of Things), and critical-path algorithms and KPIs over EPC data.

Associated Publications

The work was published in journals such as Engineering Applications of Artificial Intelligence and Journal of Zhejiang University – Science A, and presented at the IFAC World Congress 2011, IFAC-IMS 2010 and MITIP, among others. It also produced the AIRE patent (P201030790). See publications for the full list.

Reference Implementation

MAS-DUO is an open-source Python re-implementation of the thesis concepts on top of PettingZoo AEC. It is work in progress: it models the BDI–MDP–IS layers and the Ciudad Real scenario configuration, but it does not yet reproduce the quantitative results of the thesis. See the MAS-DUO implementation notes for the current status.

Thesis Supervisors and Institution

  • Supervisors: Dr. Andrés García Higuera (UCLM) and Dr. Javier Otamendi Fernández de la Puebla
  • Tutor: Dr. José Bravo Rodríguez (UCLM)
  • Institution: Universidad de Castilla-La Mancha, Ciudad Real, Spain
  • International mention: Institute for Manufacturing (IfM), University of Cambridge, United Kingdom
  • Defence date: June 2012
  • Distinction: Cum Laude, International Doctor