Distributed Models for Decision Support Jose Cuena Sascha

Distributed Models for Decision Support Jose Cuena & Sascha Ossowski Pesented by: Gal Moshitch & Rica Gonen

Motivation l The outside world is full of systems which are governed by complex laws of behavior. l Those systems can be: – Unanimated entities governed by laws of physics – Organizations of humans with artificial process rules. l Often there is a need to influence their dynamics into a desired direction.

Motivation l l For example: – Computer networks - managed in order to maintain upper bound on message delays. – Road traffic flows - influenced to avoid traffic jams. – Air traffic control – influence the planes’ routes to avoid accidents. The goal: – Maximize efficiency. – Minimize negative impact of faults.

Motivation Increasing data volume Decreasing time horizon Computer applications support the responsible person = Decision support systems (DSS)

Outline We will discuss: 1. Construction principles of DSS. 2. Distributed AI (DAI) models and architectures that applied in DSS. 3. Applications for energy management and traffic management.

Construction Principles of DSS. l Modeling DSS: – A set of world states S l given by values of the state and control variables. – ideal states, undesired states, l State values and control variables that should be achieved or avoided.

Construction Principles of DSS. l Modeling DSS: (cont. ) – A notion of preference on states l “How close” one state is to another - Partial order/Metric – A set of control actions l Control variables are changed directly l State variables are modified indirectly during the system evolution

Construction Principles of DSS. l Crucial questions DSS should know the answer on: – What is happening? l “understand” a situation by identifying advantageous and problematic aspects. – What may happen? l The evolution of the system if no intervention takes place. – What should be done? l Which are the most convenient actions improve the results.

Knowledge-Based DSS l Knowledge. Based DSS apply - divide and conquer strategy. l An example of a task-methodssubtasks tree (TMST).

Knowledge-Based DSS l Task-oriented modeling: – The classification task l l – The diagnosis task l – classifies the situation with respect to its desirability. Output set of problematic features of the current situation. An explanation that identifies the causes of such undesirable behavior. The prediction task l Evaluates how state S will evolve into state S’ given certain values for the control variables.

Knowledge-Based DSS l Task-oriented modeling: (cont. ) – The option generation task l – Generates a set of plans to overcome the problems identified. The action selection task l Selects which of the potential plans will be the outcome of the management process.

Distributed AI (DAI) Models l l Agent-based structuring introduces a more complex notion of modularity to computer science. Notion of agents allows: – Level of specialty l – Level of autonomy l l Designing agents that specialized in basic functions Integrate in an agent a set of functions required for the whole application but limited in scope. (i. e. time, space). Generality of agent allows: – Human principles for structuring organizations as design criteria.

Distributed AI (DAI) Models l l l The coordination problem has two solutions approach: Centralized – Special coordinator agent responsible for detecting interdependencies between the local agents’ activities. Decentralized – No such special agent exists – Agents interact laterally – They have the knowledge to discover inconsistencies between their intended actions and mutually adapt their local decisions.

Distributed AI (DAI) Models

Distributed AI (DAI) Models l Centralized approach: – All possible cases of inconsistencies analyzed a priori and taken into account by upper level modules. – Disadvantage l If additional lower level models are introduced, a sequence of changes has to be produced in the upper level models.

Distributed AI (DAI) Models l Decentralized approach: – Advantages: l l – Systems that are easier to build (defined very accurately only at the local level) Easy maintenance Stable coexistence independent of the number of agents in society. No problems of propagation to upper levels appear. Disadvantage: l Quality of the intelligence of the whole society of agents.

Distributed AI (DAI) Architectures for DSS. l The architecture does not consider computation and efficiency. l It considers only features necessary for different case studies.

Distributed AI (DAI) Architectures for DSS. l The architecture is built around three major components: – A perception subsystem l – An intelligence subsystem l – Allows the agent to be situated in the environment and in society by perceiving agent messages. Manages the different aspects of information processing as well as individual and social problem-solving. An action subsystem l l l Enacts the plans produced by the intelligent subsystem Displaying messages to the control personal Sending messages to other agents

Distributed AI (DAI) Architectures for DSS. l The architecture is composed of three models: – Information Model – Knowledge Model – Control Model l Information model and knowledge model focuses on the intelligence subsystem l Control model focuses on the action subsystem.

Information Model l The agents’ dynamic beliefs about the world itself and the others are stored in the information model. l The perception subsystem writes data on the information model. l When the intelligence subsystem’s knowledge is enacted, the information model is modified. l The action subsystem reads from the information model.

Information Model l The information model composed of two types of information: – Problem-solving information l Local problem-solving tasks information l Social problem-solving task information – Control information l An agenda of what is “intended to be done” – Task agenda – keeps track of the tasks to be achieved locally. – Conversation agenda in which it is involved. – keeps track of the social methods

Knowledge Model l Agent knowledge can be classified from two perspectives. – Problem solving knowledge l – Strategic knowledge l l which actions to take helps to choose among different options that the intelligence subsystem is to process next. Agent knowledge can be classified according to its role – Individual agent knowledge – Social knowledge

Knowledge Model l Individual agent knowledge: – Motivation knowledge l – A collection of patterns modeling different classes of events considered by the agent as relevant in the external world. Local problem-solving knowledge l l Basic methods – perform elementary functions by specific algorithms or constraints. Compound methods – TMST tree, rules or hard-coded simple algorithm.

Knowledge Model l Individual agent knowledge (cont. ): – Local strategic knowledge l Generation of the TMST tree. l At every level and for every task selects the method to be used.

Knowledge Model l Social knowledge: – Acquaintance models l l – Knowledge about other agents is stored in these models. By application of a pattern matching method it can be deduced whether and up to which degree some acquaintance provides desired characteristics. Social strategic knowledge l l Determines the next conversation to work on. Generation of the TMST tree when methods of different agents integrated.

Knowledge Model l Social knowledge (cont. ): – Social methods: l Copes with a task by solving its subtasks l Specify at a very high level how these subtasks are to be integrated. l Task assignment – Selection of an agent, when several available.

Knowledge Model l Social knowledge (cont. ): – Social methods (cont. ): l Task synchronization – Once tasks are assigned, the flow of information between them needs to be synchronized. l Solution – integration The results of subtasks of a social method need to be adapt to each other in order to receive a consisting result.

Control Model Perception Subsystem Intelligence Subsystem Strategic Knowledge Motivation Knowledge Local Strategic Knowledge Acquaintance Models Social Strategic Knowledge Action Subsystem Information Model Messages Perceptions Local Problem Social Problem Information Solving Task Agenda Conversation Agenda Problem Solving Knowledge Local Social Methods Social Problem Solving Inf Messages Control Inf Actions

Control Model Perception Subsystem Intelligence Subsystem Strategic Knowledge Motivation Knowledge Local Strategic Knowledge Acquaintance Models Social Strategic Knowledge Action Subsystem Information Model Messages Perceptions Local Problem Social Problem Information Solving Task Agenda (Add) Conversation Agenda Problem Solving Knowledge Local Social Methods Social Problem Solving Inf Messages Control Inf Actions

Control Model Perception Subsystem Intelligence Subsystem Strategic Knowledge Motivation Knowledge Local Strategic Knowledge Acquaintance Models Social Strategic Knowledge Action Subsystem Information Model Messages Perceptions Local Problem Social Problem Information Solving Task Agenda (Execute Sum) Conversation Agenda Problem Solving Knowledge Local Social Methods Social Problem Solving Inf Messages Control Inf Actions

Control Model Perception Subsystem Intelligence Subsystem Strategic Knowledge Motivation Knowledge Local Strategic Knowledge Acquaintance Models Social Strategic Knowledge Action Subsystem Information Model Messages Perceptions Local Problem Social Problem Information Solving Task Agenda Conversation Agenda Problem Solving Knowledge Local Social Methods Social Problem Solving Inf Messages Control Inf Actions