1 Properties of agents: Properties of agents: rationality - PDF document

2. Single-agent systems 2.1 Formal Definitions and Properties n Term “agent” used with many different meanings: Our goal: rather broad definition l Agent of an agency (see James Bond) l Agent as a person acting on behalf of another Definition : Agent person An agent Ag is a triple Ag = ( Sit , Act , Dat ) where l Agent as a substance achieving some n Sit is a set of situations Ag can be in change/effect n Act is a set of actions Ag can perform n Even in Computer Science no agreement! n Dat , Ag ’s internal data, is the set of possible values n Also under discussion if there is such a thing as a Ag ’s internal data areas can have single-agent system (agent acts in environment and From outside, Ag seems to be a “function” g Ag : Sit ‘ Act , some see environment as a second agent) but in reality we have f Ag : Sit 5 Dat ‘ Act Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger Remarks (I) Remarks (II) n g Ag is not really a function, it can be a relation! n Our definition is really broad: l If we want to use observations (from outside) to l Every computer program fulfills it. predict the behavior of an agent, we can be in l Human beings trouble! l Animals l Even if we know the real function f Ag ,predictions l Lifeless objects are not easy if we do not exactly know the actual F Some authors require additional properties from an values of the agent’s internal data areas. agent, l Many problems that we have with multi-agent but different authors require different properties and systems have their roots in this problem! some of them cannot be easily formalized! Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger Properties of agents: Properties of agents: autonomy knowledge-based cogito ergo sum An agent is autonomous , if it tries to achieve goals on its An agent acts knowledge-based , if it can justify its actions own without direction from other agents. An autonomous agent can react even when encountering unforeseen events. Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger 1

Properties of agents: Properties of agents: rationality reflectivity and reactivity (I) $ $ $ ? $ $ $ $ An agent is and acts rational , if its behavior is oriented An agent acts reflective , if it bases its decisions on how to on fulfilling its goals as completely and as good as act very much on the actual value from Dat . possible. Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger Properties of agents: Properties of agents: reflectivity and reactivity (II) social No ! Else it acts reactive . An agent acts social , if it does not hinder other agents in their actions Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger Some formal properties of agents: action language of an agent Remarks bbaa abaa n Does this mean an agent can “do” only one thing at a bba aba time? bb ab No! b a We model performing actions a 1 , a 2 , and a 3 at the same time by performing action a 1,2,3 that is defined as performing a 1 , a 2 , and a 3 at the same time. n In most cases, we are interested in sequences of Let Ag = ( Sit , Act , Dat ) be an agent. A set L Ag Õ Act * is an actions of agents (i.e words of the action language), action language to Ag , if for each word w Œ L Ag holds: not only single actions/decisions. each start word v of w (i.e. w =vu) is element of L Ag Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger 2

Some formal properties of agents: Some formal properties of agents: deadlock free liveliness S O S ! ? Let Ag = ( Sit , Act , Dat ) be an agent with action language An agent Ag = ( Sit , Act , Dat ) with action language L ag is L Ag . An action t Œ Act is alive in L Ag , if for all w Œ L Ag called deadlock free , if there is no w Œ L Ag such that for holds: there is a v Œ Act * with wvt Œ L ag . all v Œ Act *holds: wv œ L Ag . Ag is called alive , if for all t Œ Act t is alive in L Ag . Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger Some formal properties of agents: fairness Remarks n The properties described first obviously are of a certain relevance when looking at multi-agent systems. n The formal properties will be needed later to show a very basic problem of multi-agent systems: Let Ag = ( Sit , Act , Dat ) be an agent with action language Even if all agents in a MAS have a certain property, L Ag . Ag is fair , if for all infinite words w with all start then the MAS as a whole is not guaranteed to have words of it within L Ag and all t Œ Act holds: this property. If there are infinitely many start words v i of w, such that v i t Œ L Ag , then t appears in w infinitely often. So, do not immediately forget these properties. Multi-Agent Systems Jörg Denzinger Multi-Agent Systems Jörg Denzinger 3