Beruflich Dokumente
Kultur Dokumente
Towards intelligent decision support systems for emergency managers: the IDA approach Adam Maria Gadomski, Sandro Bologna and Giovanni Di Costanzo
ENEA Italian National Research Agency for Energy and the Environment, C.R. Casaccia, 00060 Rome, Italy Fax: +39 06 3048 6511 E-mail: gadomski_a@casaccia.enea.it
Anna Perini
IRST-ITC Istituto Trentino di Cultura, I-38050 Povo TN, Italy Fax: +39 461 302040 E-mail: perini@itc.it
Marco Schaerf
Universit di Roma La Sapienza, Via Salaria 113, 00184 Rome, Italy Fax: +39 6 85300849 E-mail: schaerf@dis.uniroma1.it
Abstract: The paper presents ENEAs next step towards the development of Intelligent Decision Support Systems (IDSS) for large-scale industrial and territorial emergencies. The prototype IDA (Intelligent Decision Advisor) for emergency management in an oil port is analysed as a test case. The work was performed under the national R&D MICA project and specifically ENEAs long-term strategic MINDES Program synchronized with indications of the worldwide GEMINI (Global Emergency Management Information Network Initiative) of the G7 Committee. IDA is an approach in designing intelligentagent based kernels of IDSS. In the frame of the generic TOGA (Top-down Object-based Goal-oriented Approach) model of abstract intelligent agents, IPK (Information, Preferences, Knowledge) architecture was employed. The specific IDA objectives were to develop and verify the properties of an information-managed agent and a knowledge managed agent, where the latter should suggest an action or plan after every new significant event in the emergency domain. The IDA functional kernel is composed of three simple agents: a DirectAdvisor, which interacts with the human user and emergency domain, an InfoProvider, which manages information and intervention goals and an IDAPlanner, which plans adequate interventions. For the design, UML (Unified Modelling Language) has been employed. MDP (Markov Decision Process) and CBR (Case-Based Reasoning) are used for planning crisis management actions. Owing to a generic agent model and object-based conceptualization, the IDA system should be adaptable to the different roles of emergency managers. The obtained results confirm the IPK conceptualization hypothesis and provide a concrete technological experience for the next step towards high-intelligent DSSs for the management of emergencies. Keywords: Intelligent Decision Support System; emergency management; abstract intelligent agent; planning. Reference to this paper should be made as follows: Gadomski, A.M., Bologna, S., Di Costanzo, G., Perini, A. and Schaerf, M. (2001) Towards intelligent Copyright 2001 Inderscience Enterprises Ltd.
225
226
Introduction
New things get started by evolution of chance, not design Allan Newell
The first prototype of an Intelligent Decision Advisor (IDA) for emergency management has been the objective of the national R&D project MICA 2.8.3.1.F focused on the development of an active computer support for large-scale industrial and territorial emergencies. The project has been realized under ENEAs long-term strategic MINDES Program synchronized with the worldwide GEMINI (Global Emergency Management Information Network Initiative) of the G7 Committee [1,2]. The main objective of MINDES (Managerial Intelligent Node for Decisional Emergency Support) is to develop an intelligent decision-support kernel for the computerized nodes of an Emergency Management Network. Local decision support systems should be connected by the internet with other similar emergency management centres in the structure of an international Global Emergency Management Information Network. The accepted software solutions are based on intelligent agent theories and technology [35]. The MINDES Program profits from the past experience of ENEA in the field of emergency management and plant-operator support systems. Most importantly, the aim of the program is to reduce the probability of human managerial errors during decision making. The following strategic objectives of MINDES were defined: To provide real-time data necessary for the emergency managers decision making during emergency situations. To use data from the available information systems present in and out of the test site. The end-user test site could be, for instance, a chemical plant system (oil refinery) or a regional emergency management organization. To have a user friendly interface to assist high level industrial or administration managers. An easy to use communication interface will be supported by multimedia techniques, a GIS system, and a voice commands option. Therefore, the use of the system should not require any help from computer specialists.
It should be a tool for periodic manager training sessions on emergency games. In the above context, the IDA contribution to the MINDES-GEMINI program was defined as a one-year R&D project performed with the contribution of IRST Trento and the University La Sapienza of Rome. The project was focused on the modelling of the domain of emergency and a verification and validation of some intelligent agent technologies in a selected class of concrete emergency management cases. The paper deals with three subjects related to ENEAs experiences and ongoing results: Motivation for the development of intelligent multipurpose DSS and IDA objectives. Theoretical and methodological frameworks. IDA; a prototype of IDSS.
227
Theoretical background
228
Figure 1
Here [3,9], the information concept represents every state of the pre-selected intervention domain, knowledge is every mental/abstract entity/system which is able to transform information into other information and preferences are relative rules which order domain states according to a subjective importance scale accepted by a human domain expert, (for instance, using certain utility/value function). Every IPK triangle, called a monad, considers preferences or knowledge base, on a lower meta-level as its external domain of activity. A simplified repetitive mechanism of a personoid is the following: New Data from the Domain modifies Information. Information activates Preferences. Preferences produce a Goal. Goal activates Knowledge. Knowledge modifies Information.
If Preferences or Knowledge are not able to execute such a task, then it produces Data for the higher-level triangle and itself becomes the new Domain, for instance, for a planning function. In this way, a personoid has many possible abstract domains of activity on different abstraction/meta levels of reasoning. They are: an image of the real emergency domain, agent preferences bases and agent knowledge bases. The concepts: intervention domain,
229
abstract intelligent agent, information, preferences and knowledge consist of a basic ontological platform of the TOGA theory. In such a context, the main IDA objectives [10] were to develop and verify information management functions by the InfoProvider agent, and knowledge management by the Planner agent, where the latter should suggest an action or plan (a sequence of actions) after every new significant event in the emergency domain. Because of numerous technical problems that ought to be solved, the PreferenceManager agent has been substituted by the fixed set of the possible goals of intervention. Its modelling and realization still require a theoretical analysis. It should be added to the system in the IDA-2 version. From a technical point of view, the work has been concentrated on: Modelling and formalization of emergency domains, in particular, on their representation by a generic world of objects with events, resources and facts. The dynamics of the domain has to be introduced by qualitative relations between components of the domain. Choice, modelling and verification of planning methods and their implementation.
Figure 2 illustrates the functional decomposition of the personoid on the prototypal IDA agents.
Figure 2 Decomposition of the personoid IPK structure into the three IDA agents for the verification of information management and planning functions
230
B 4
The emergency management top-goal is to moderate crisis events preventing accident degeneration and minimizing total losses, i.e. the initial emergency state has to be changed to the state accepted by the port manager as well as the risk generation process caused by the fire of one of the six tanks, stopped. The local emergency manager must use his/her own resources in the best way; he/she has a set of operative units at his/her disposal with the help of which everyone can execute a predefined set of tasks.
231
3.2 Descriptive and operational knowledge 3.2.1 Descriptive knowledge: domain model
In this section we discuss how domain dependent information is transformed. In particular, we will describe the state and the action representation used for the system planning function. A generic emergency domain model referring to some class of the emergency domains is considered as a descriptive domain-knowledge; formally it works as follows: KD j : Ii Ii+1, for i=1,2,, where: Ii I i+1 are two informations which are components of a specific description of an emergency domain state, or, in general, of the domain of activity of a monad, KDj is a domain descriptive knowledge, which is part of the relations among classes of abstract objects and their attributes present in the domain model. Every concrete element of the domain is represented as an instance domain object or resource. For example, domain objects classes of the oil port are tank, tanker, dock and rack. The system states are described by a finite set of state variables (attributes of the domain objects) that take on discrete values; they model features of the object domains. For instance, the temperature of a tank is modelled by a Boolean variable that states if the object temperature is above or below a critical value or the level of coverage by foam of an object is an integer ranging from 0 (no foam) to 3 (fully covered). Table 1 illustrates the classes of state variables (attributes) currently used for the oil port domain.
Table 1 IDA descriptive domain knowledge in the form of a table Dock irradied spilled temperature foamCover fireRisk eventSize defiled WharfAccess Tank irradied spilled temperature foamCover fireRisk eventSize fireTop productLevel Tanker irradied spilled temperature foamCover fireRisk eventSize productLevel tankerAtWharf inDock Rack irradied spilled temperature foamCover fireRisk eventSize
Type / Object 4 B B 4 B 10 B 4 B B B B
Note: State variables are defined for the domain objects. The first column (type) indicates if the variable is Boolean (B) or the number of discrete values that it can assume
So for a given set of variables (object attributes) used to describe states of the environment, for instance {Tank1.irradied, Tank1.temperature, Tank1.foamCover}, their specific state (information) can be s={Tank1.irradied=3, Tank1.temperature=0,
232
Tank1.foamCover=0}. In this way the descriptive knowledge can be enriched simply by adding new state variables to the domain model. This operation can be performed by the system user.
Towards intelligent decision support systems for emergency managers the class of means/resources that can be employed;
233
a qualitative estimation of the action cost (for example 0=null, 1=low, 2=medium, 3=high).
The IDA prototype is a semi-intelligent agent with the capability to execute a certain class of tasks autonomously [4]. It is composed of the following functional agents: DirectAdvisor, InfoProvider, IDAPlanner.
Altogether, they have the capability to: represent the current emergency domain state and update it while new events are signalled by the InfoProvider agent; represent user goals notified by the InfoProvider agent; suggest an action or a plan to the user by DirectAdvisor; forecast the environment state upon request by DirectAdvisor or upon execution of a given action.
The IDA architecture signalled in Figure 2 has been designed and implemented with the support of the Rational Rose Visual Modelling tool using UML (Unified Modelling Language) notation [14]. It will be illustrated in more detail in the following.
4.1 InfoProvider
The InfoProvider agent is a common interface for other agents and for the IDA user. The main tasks performed by the InfoProvider can be divided into two groups, since we pointed out two main functionalities in the IDA system. First of all, the InfoProvider provides current information to the two other agents on the actual state of the emergency domain and resources. These data are stored in a relational database (using DBMS Ms Access 97). It manages all domain dependent information and provides it to the other agents upon request; it acts as a mediator for the Planner and for the Direct Advisor. All information related to an emergency domain map to an abstract domain of activity for the IDA. The second group of functions is related to the updating and management of the database upon request. The InfoProvider has the capability to reason on these data in order to provide information at a higher level of abstraction. The database contains information, goals and the user operational knowledge in the following relational tables: Objects ObjectState Resources Resource State domain object list divided into fixed classes; values of the object attributes at any given moment; resource object lists; values of the resource attributes at any given moment;
234
A.M. Gadomski, S. Bologna, G. Di Costanzo, A. Perini and M. Schaerf plausible goals of local intervention; the set of actions (operational knowledge) appropriate to the selected roles of emergency manager.
New components of these tables can be added at run-time. The database can also be seen as a threefold role component, i.e. as: 1 2 Interface to other modules. Abstract conceptual components: facts, actions and goals. The abstract components contain the knowledge about use of the resources and the evolution of the emergency (historical data). Actual components: instances of objects and resources. They represent the current state of the domain.
The analytical results obtained here suggest that the temporal intervals, non-monotonic and default reasoning [1517] could also effectively support a generic inference tool employed for information, knowledge and preference management in the frame of the IPK architecture.
4.2 IDAPlanner
Emergency management or crisis mitigation planning problems requires to be modelled as an interactive decisional process where the human and the machine reasoning activities interleave. For instance, given an emergency situation such as a fire on a tanker at an oil port dock, the following interactions between the human and the decision advisor system can be described: the human decision maker can pose specific goals devoted to maintain under control/surveillance a critical zone of the port, such as neighbouring tanks, in order to avoid the evolution of the emergency into a catastrophe. So, for instance he/she can express goals such as maintaining the temperature of neighbouring tanks under critical values or reducing the amount of inflammable liquids near the crisis epicentre; the system can suggest appropriate actions/tasks to the human decision maker to be performed in order to manage the emergency situation, taking into account the specific intervention goals and, in parallel, evaluating action costs versus their benefits [18]. So, for instance, the system can suggest taking the tanker away from the dock and letting it burn instead of trying to extinguish the fire at the dock, thus maintaining a high risk of fire propagation to the other port resources; the system can be asked to forecast the state of the oil port after the execution of a given action; the decision maker decides which action to perform.
This kind of planning problem is also known as Mixed-initiative planning [19] and the IDAPlanner agent was designed according to its basic requirements. They have the ability to: represent the current state of the environment (for instance the oil port) and update it while new events are notified by the InfoProvider agent;
Towards intelligent decision support systems for emergency managers represent user goals notified by the InfoProvider agent; suggest to the DirectAdvisor an action or a sequence of actions; forecast to the DirectAdvisor the environment state upon execution of a given action.
235
Moreover the IDAPlanner agent was built according to the general requirements of system extensibility and of applicability to different concrete domains. In particular: the requirement of realizing the agent competencies , (the planner meta-knowledge and meta-preferences) using different approaches ranging from deterministic to nondeterministic AI planning approaches; the requirement of building a planner whose domain model could be improved or extended or substituted with a different emergency domain.
The first requirement was met by defining the IDA class hierarchy (a C++ class) depicted in Figure 4. The IDAPlanner class is implemented as a complex planner module and it realizes the IDAPlanner competencies described above.
Figure 4 The IDA class diagram with the MDP Planner class specialization, UML notation [14]
236
The classes that specialize in the IDAPlanner class, represent different ways of its possible realization. In the IDA system, a decision-theoretic scheme based on the Markov Decision Processes (MDP) techniques [20] is employed to the planning process, and the MDP Planner class has been defined. Dependent on the future applications of IDA, other planning methods can be applied for these specializations. The second requirement has been taken into account by isolating domain dependent information from operational knowledge (available actions) into the action library. In this way, a functional module that is domain and situation dependent can be extended or replaced by another. Both requirements have been inserted as the properties of the MDP Planner class and implemented as the MDP Planner agent, a current realization of IDAPlanner.
237
More detailed information related to the architecture of the IDA Planner is included in the report [24]. Modelling an intervention planning with respect to the MDP framework also posed some interesting problems such as how to model complex actions with various duration times and actions that can be executed in parallel [15]. Concluding, from the personoid perspective, the planning process is realized on the second and third meta-levels. It can use different criteria (meta-meta-meta-preferences) for internal choices and, in practice, can be realizable by different planning and learning methods. These criteria were identified during the project and could be inserted as an independent meta-preference base for planning, and meta-meta-preferences for learning management, in the next IDA version.
The objective of the DirectAdvisor is to be an interface among information sources, i.e. InfoProvider and intervention Planner, and human manager. The user has three types of interactions with IDA: 1 Map set-up session, when various domain maps are inserted.
238 2 3
A.M. Gadomski, S. Bologna, G. Di Costanzo, A. Perini and M. Schaerf Emergency Set-up session when the type of emergency, initial emergency situation, possible intervention goals and the managerial role are edited. IDA Demo Emergency-Management Interactive session.
The Demo illustrates a scenario of a simple industrial emergency, a kind of emergency game. From the perspective of game theory it can be called a game with nature. The domain and rules of this game can be modified/updated by the player, using the interface module of the DirectAdvisor. The player is an emergency manager with a predefined role. Here, we illustrate how, by an interactive generic conversation scenario between IDA and its user, IDA may help make decisions. Let us now assume a play convention: He/she plays with a simulated emergency. He/she tries to achieve emergency management goals.
The player activities are focused on: current emergency state identification cause searching (backward-propagation) consequence searching (forward-propagation) objective searching (current max. preferred state) in preferences base elaboration of a suggested intervention (an action), as a plan, instruction, task.
In general, we have player decisions: Without support; where the player has to choose data sources, data/information, action or to elaborate and to execute actions/intervention himself. With support; where the player needs to choose only among actions suggested by the system.
In the current IDA version the map, user role and initial emergency situation are preloaded. The generic IDA scenario is as follows: 1 2 3 4 E-Mer (emergency manager) inserts new information (facts) about a change in the emergency domain. IDA provides data about the current state of emergency domain (menu-driven). E-Mer chooses intervention-goal from the possible goals list. IDA presents intervention-plan as a sequence of actions.
239
E-Mer inserts new facts; for instance, when the objective of the suggested action was not achieved the user should decide: go to 3 or go to 4.
240
Conclusions
IDA is a prototypal, intermediate, demo system focused on the validation of pre-selected properties of a multipurpose IDSS. It has been constructed taking into consideration the general framework and paradigms of the IPK architecture, and the personoid framework, as a specific abstract intelligent agent. From the theoretical point of view, major attention in this first version of IDA has been concentrated on the representation of the emergencydomain descriptive knowledge as the TOGA ontology [25] of the abstract-objects world, and on its concrete modelling using Object-Oriented specification languages [14]. The operational domain knowledge has been conceptualized as STRIP- like operator. A planning method, considered as meta-knowledge, with implicitly included knowledge meta- and meta-meta-preferences, has been defined. It was realized in the IDA system by exploiting the Markov Decision Process techniques. The obtained results should enable structural separation of preferences and knowledge, on the different IDA reasoning levels. The hypothesis of the IPK conceptualization framework has been practically verified through the modelling of an emergency in an oil port. Therefore, in the next version of IDA, we expect to implement a Preference Management Agent, with an automatic choice of intervention goal (suggested to the manager), and to separate information, knowledge and preferences on the systems structural meta-levels. This separation will enable the auto-modification of employed reasoning methods, preference rule bases, and operational knowledge available for the manager. However, the realized prototype satisfied our expectations related to the current research phase of ENEAs MINDES Program. The currently adopted ontology models (descriptive knowledge) and software solutions, such as domain independent learning and planning methods, should also be tested and reused in other more applicative projects focused on Intelligent Decision Support Systems for various particular emergency domains. For example, one of the most critical factors of the managerial real-world decision support is the necessity of decision making under uncertainty and incompleteness of available information. In the case of the personoid model, the appropriate methods could be added, modified and validated without modification of the invariable, repetitive, incremental and recursive IPK architecture of the IDSS kernel.
Acknowledgement
We need to mention the colleagues who also participated in the different phases of the developing of the IDA system, they were: F. Ricci, C. Balducelli, L. Senter, R. Sensi and R. Iannucci. We are very grateful for their valid contributions.
241
5 6
10
11 12 13
14 15 16 17 18
19
20
Gadomski, A.M. (1994) TOGA: a methodological and conceptual pattern for modelling of abstract intelligent agent, Abstract Intelligent Agent,1993 , A. M. Gadomski (Ed.) Published by ENEA, Rome, February, pp, 1525. Franklin, S. and Graesser, A. (1997) Is it an agent, or just a program?: a taxonomy for autonomous agent, In Proceedings of the Third International Workshop on Agent Theories, Architectures, and Languages, published as Intelligent Agents III, Springer-Verlag, pp.2135. Huhns, M.N. and Singh, M.P. (1998) (Eds.), Readings in Agents Morgan Kaufman Publishers, San Francisco. Gadomski, A.M., Bologna, S. and Di Costanzo, G. (1995) Intelligent decision support for cooperating emergency managers: the TOGA based conceptualization framework, In Proceedings of TIEMEC 1995: The International Emergency Management and Engineering Conference, J.D. Sullivan, J.L. Wybo and L. Buisson (Eds.), TIEMS Press, Nice, pp.379385. Balducelli, C., Bologna, S., Di Costanzo, G., Gadomski, A.M. and Vicoli, G. (1995) Computer aided training for cooperating emergency managers: some results of the MUSTER project, In Proceedings of the MEMbrain Conference on International Aspects of Emergency Management and Environmental Technology, H. Drager (Ed.), A/Q Quasar Consultants Press, Norway, Oslo, pp.433444. Gadomski, A.M., Balducelli, C., Bologna, S. and Di Costanzo, G. (1998) Integrated parellel bottom-up and top-down approach to the development of agent-based intelligent DSSs for emergency management, In Proceedings of the International Emergency Management Society Conference TIEMS98: Disaster and Emergency Management , J.R. Harrald and G.L. Shaw (Eds.), TIEMS Press, Washington, pp.421434. Gadomski, A.M. and Di Costanzo, G. (1996) Intelligent decision support system for industrial accident management , In Proceedings of 8th European Simulation Symposium, A.G. Bruzzone and J.H. Kerckhoffs (Eds.), SCS International Press, Genoa, pp.137141. Balducelli, C., Di Constanzo, G., Gadomski, A.M. and Iannucci, R. (1997) Requisiti funzionali per un sistema intelligente di supporto alle decisioni nel campo della gestione delle emergenze, The ENEAs Internal Report, Rome. Dean, T. and Wellman, M. (1991) Planning and Control, Morgan Kaufman Publishers, San Mateo. Veloso, M. (1996) In Advanced Planning Technology, A. Tate (Ed.), AAAI Press, May, pp.277282. Aamodt, A. and Plaza, E. (1994) Case-based reasoning: foundational issues, methodological variations and system approaches, Artificial Intelligence Communications, J., IOS Press, Vol. 7, No. 1, pp.3959. Quatrani, T. (1998) Visual Modelling with Rational Rose and UML. Addison-Wesley Longman Inc, Reading, Massachusetts. Allen, J.F. (1983) Maintaining knowledge about temporal intervals, Communications of the ACM, November, Vol. 26, No. 11, pp.832843. McCarthy, J. (1980) Circumscription a form of non-monotonic reasoning, Artificial Intelligence, Vol. 13, pp.2739. Reiter, R. (1980) A logic for default reasoning, Artificial Intelligence, Vol. 13, pp.81132. In the frame of the IPK architecture, the evaluation of the action costs, i.e. the application of choice criteria to operational knowledge, can be represented as an activation of a preference rule base on the first meta-level of the personoid structure, see Figure 2. Ferguson, G., Allen, J.F. and Miller, B. (1996) TRAINS-95: towards a mixed-initiative planning assistant, In Proceedings of the Third International Conference on AI Planning Systems (AIPS-96), B. Drabble (Ed.), Edinburgh, Scotland. Puterman, M. (1994) Markov Decision Processes, John Wiley and Sons.
242
21
22
23
24
25