Valoraciones de la Restauración

Sadri et al. present in [STT01] a formal, logic-based approach to one-to-one agent negotiation, in the context of goal achievement in systems of agents with limited resource availability. Agents

have a dichotomous (0-1) cardinal utility function as in this thesis, such that an agent’s utility value of a bundle is 1 if that bundle allows the agent to achieve its goal, 0 otherwise. More specifically, each agent is equipped with a knowledge base, which would contain beliefs such as the resources held by the agent, as well as the goal the agent is equipped with, an intention that might consist of the goal, a plan to achieve that goal, which may be to obtain a certain bundle of resources, and an associated set of available and missing resources that are needed to carry out the plan.

The purpose of negotiation in [STT01] is for agents to obtain missing resources, while retaining available ones that are necessary for the plan in its current intention. A dialogue move (or utterance) in [STT01], and somewhat similarly in this thesis, is an instance of a schema tell(X, Y, Subject, T), where X is the utterer and Y is the receiver of the dialogue move, and T is the time when the dialogue move is uttered. Subject is the content of the dialogue move, expressed in some given content language. A language for negotiation in [STT01] is a set of dialogue moves and, as part of a given language, there are two subsets of dialogue moves: initial moves and final moves. An example of an initial move could be a request that is initiated by an agent for a resource that it requires. Each final move is either successful or unsuccessful. Examples of the types of dialogue moves that may make up a language are resource requests, acceptance of moves, refusal of moves and promise to exchange resources (see [STT01] for a possible concrete language for negotiation).

Conversation between agents in [STT01], and similarly in this thesis, is based on dialogues, where a single dialogue is a sequence of dialogue moves (utterances), according to a pre-defined acceptable sequence of utterances (the protocol), used to obtain a resource and such that agents alternate utterances in a dialogue. Intuitively, a dialogue should begin with an initial move, according to the given language for negotiation. Agent dialogues can be connected within sequences (of dialogues), all aimed at achieving an individual agent’s goal. Sets of sequences of dialogues aim at allowing all agents in the system to achieve their goals, and can be used to solve the resource reallocation problem.

2.2. Dialogues in Multi-agent Systems 33 resources) is based on fully distributed agent dialogues as a way to request resources, propose resource exchanges, suggest alternative resources, and so on. The negotiation framework pre- sented is a general one for representing policies governed by protocols and generating dialogues accordingly. A negotiation policy for an agent is specified as a set of dialogue contraints, which are if-then rules of the form p(T ) ∧ C ⇒ ˆp(T + 1) where p(T ) and ˆp(T + 1) are shorthand for dialogue moves of the form tell(X, Y, Subject, T) as above and C is a conjunction of literals in the language of the knowledge base of the agent. If this is a dialogue constraint for an agent x, for example, then it is interpreted as follows: if at a certain time T in a dialogue some other agent y utters to x a dialogue move p(T ), then the corresponding instance of the dialogue constraint is triggered and, if the condition C is entailed by the knowledge base of x, then x will utter a dialogue move ˆp(T + 1), with y as receiver, at the next time T + 1. An agent program (set of dialogue constraints associated with an agent) should ideally guarantee that the agent’s utterances are conformant to the protocols governing its dialogues and that the agent’s dialogues terminate. Other ideal properties desired of an agent program are discussed by Sadri et al. [STT01] and Endriss et al. [EMST03], e.g. the agent always produces exactly one reply to a (non final) move made by the other agent.

Our communication rules in this thesis for representing agent negotiation policies combine ideas from the dialogue constraints of Sadri et al. [STT01], which do not say how agents’ beliefs are modified, and theactions of Rueda et al. [RGS02]. An action (ability) A in [RGS02] is an ordered triple < P, X, C > denoted as {X1, . . . , Xn}

A

← {P1, . . . , Pm}, {not C1, . . . , not Ck},

where P is a set of m literals representing preconditions for A, C is a set of k constraints of the form not L (L is a literal) and X is a set of n literals representing consequences of executing A. An agent, according to [RGS02], can consider an action only if its beliefs tell it that the preconditions and constraints are satisfied. The consequences specify the revisions to be made to the agent’s set of beliefs as a result of applying the action, where revision consists of removing any literal in the belief base of the agent that is complimentary of any literal in X and then adding X to the resulting set.

It is assumed in [STT01] that all dialogues between an agent i and any other agents are atomic, i.e., it is not possible for i to be engaged in two different dialogues at the same time. This is

an assumption we relax in this thesis, as follows:

Conversation between agents in this thesis will be based on dialogues (sequences of utter- ances by agents) generated by the agents’ negotiation policies (sets of communication rules) and governed by protocols represented as finite-state machines. The negotiation policies will allow for an agent to be engaged in two or more different dialogues at the same time.

Moreover, we consider one-to-many agent dialogues in this thesis as well as one-to-one dialogues as in [STT01].