Intelligent Agents: Theory and Practice
Wooldridge and Jennings
- Weak notion: autonomous + social + reactive + proactive
- Strong notion: mentalistic (K/B/I/obligation) / emotional
- optional attributes: mobility (in an electronic network), veracity, benevolence, rationality
Agents as intentional systems. Intentional notions as abstraction tools for complex systems.
Second-order intentional system: has beliefs/desires/etc. about beliefes/desires/etc.
- Information attitudes: belief, knowledge
- Pro-attitudes: desire, intention, obligation, commitment, choice
Logics of knowledge (epistemic) and belief (doxastic).
- Problems with classical first-order logic:
- Syntax doesn’t allow formulas inside functions (only terms)
- Referential opacity of intentional notions (beliefs are not truth-functional)
- Modal languages (non-truth-functional modal operators)
- Possible worlds semantics (Hintikka: epistemic alternatives, Kripke)
- necessitation rule (): agent knows all tautologies
- “” + axiom K (): agent knowledge closed under logical consequence (the agent knows everything that can be proved)
- Levesque: explicit (small) & implicit (includes the logical consequences; potentially infinite) beliefs
- Konolige: deduction structures
where is a set of inference rules.
- Represent relationships between meta-language terms denoting an agent and object language terms denoting some formula.
- Problem: easily lead to inconsistency.
Pro-attitudes: goals and desires
Adapting possible world semantics has the side-effect problem: agents having a goal of the logical consequences of their goals.
Theories of agency
Logic framework that combines the various aspects of agency, showing how they are related. It should also be capable of representing dynamic aspects.
- Moore: knowledge pre-conditions for actions. He formalised a model of ability, knowledge and actions, allowing agents with incomplete information about how to achieve a goal.
- Cohen and Levesque:
- Theories of intention: pose problems that need to be achieved, non-conflicting, success tracking, believed to be possible, no believe that they won’t be done, side effects needn’t be intended by the agents.
- Logic of rational agency + derived constructs, incl. that of persistent goal (persistent goal = goal , currently isn’t true, goal won’t be dropped until is satisfied or there’s a belief it’ll never be) which helps define intention.
- Logical framework for agent theory based on beliefs, desires and intentions as primitive modalities. Model in which B/D/I-accessible worlds are branching time structures.
- Notion of realism: how beliefs about the future affect desires and intentions.
- Singh: familty of logics for representing intentions, beliefs, knowledge, know-how and communication in a branching-time framework.
- Werner: general model of agency /based upon economics, game theory, situated automata theory, situation semantics and philosophy).
- Wooldridge: family of logics to model multi-agent systems, for use in specification and verification of realistic multi-agent systems.
Based on speech act theory: cmommunication utterances are actions.
- representatives (eg. informing)
- directives (eg. requesting)
Agent communication languages:
- KQML: syntax for messages and performatives (tell, perform, reply, etc.)
- KIF: syntax for message content (first-order predicate calculus in LISP syntax).
- Deliberative architectures (physical-symbol system hypothesis)
- Transduction problem: translating the real world to symbols
- Representation/reasoning problem
- Brooks: situatedness & embodiment lead to emergence
- Situated automata (declarative language compiled down to a circuit)
- Cognitive problem solving & distributed AI
- Interface agents
- Information agents / cooperative information systems
- Believable agents