Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
An Event-ConditionAction Logic Programming Language J. J. Alferes F. Banti A. Brogi 1 Overview Event Condition Action (ECA) languages. Updates of (ECA) rules. Dynamic Logic Programs (DyLPs). Evolving Reactive Algebraic (ERA) Programs. Syntax Semantics Example of usage. Parallel with DALI. Future works. 2 Event Condition Action languages Fundamental construct: ECA rules On Event if Condition do Action When an Event occurs, under some Conditions, execute a given Action. Example On callE(Call,Emp) if office(Emp, N) do forw(Call,N). Events and Actions can be simple or complex, resulting from the combination of basic ones. 3 Updates of ECA rules. Problem: How to update the behaviour of an ECA System? Example: Suppose that, whenever an employer is in a working meeting, the phone call must not be forwarded to him. Solution: To allow the system to assert new rules and to encode exceptions to existing ECA rules. Inhibition rules: Rules that specifies when actions are not executed. Example: When callE(Call,Emp), meeting(Emp,) do not forw(Call,N). 4 Logic Programs Updates We consider sequences P1, ... , Pn (Pi) of LPs called Dynamic Logic Programs (DLPs) Basic language: generalized logic programs P1 is the initial program, while the others Pis are updates The head of a rule can be either a positive or negative literal. Common ideas: Stable model approach Causal rejection principle: A rule is rejected iff there is a more recent rule whose body is satisfied and is in conflict with the old one. (i.e. the two rules has complementary heads). 5 Syntax of ERA ECA rules: On Event if Condition do Action Data and inference rules: L. L0 ← L1, …Ln. Every Li is an atom or the negation of an atom. Inhibition rules: When B do not Action. B is a conjunction of literals and events. Event is a literal representing a simple or complex event. Complex event are obtained by an algebra of operators. Condition is a conjunction of literals. Action is an atom representing a simple or complex action. Complex actions are obtained by an algebra of operators. 6 Syntax of ERA Syntax of events: e::= eb |e1 ٧ e2| e1 ٨ e2| A(e1, e2,e3)| edef. Syntax of basic and complex actions: ab::= ax| raise(eb)| assert(r)| retract(r)| define(d). a::= ab| a1 ► a2 | a1||a2 | IF(C, a1, a2) | adef. Event and action definitions: edef is e. adef is a. 7 Semantics of ERA Events are atoms occurring in LP Ei called input programs and represent external inputs and internal events. Inference system based on the refined semantics of DyLPs. If an atom “Action” corresponding to an action is inferred, the corresponding action is executed. Replace On Event if Condition do Action with: Action ← Condition, Event. Replace When B do not Action with: not Action ← B. Replace edef is e and adef is a with: adef ← e and edef ← e. 8 Semantics of ERA(cont.) Add inference rules for deriving literals: e1 ٧ e2 ← e1. e1٧ e2 ← e2. e1 ٨ e2 ← e1 ,e2. The semantics of actions is given by a transition system describing step by step how the execution of an action affects a program. The execution of a complex action may require several steps formed by basic actions. < P, Ep, Ei.EF> → G <P’, Ep’..Ei, EF’> 9 Example On callE(Call,Emp) if office(Emp, N) do forw(Call,N). assert(r) r : When callE(Call,emp), meeting(emp) do not forw(Call,N). New behaviour: every call to emp is forwarded to his mobile phone. assert(r1) ► assert(r2). r1: When callE(Call, emp) do not forw(Call, N). r2 : On callE(Call,emp), mobile(emp,m) do forw(Call,m). Return to previous behaviour: retract(r1) ► retract(r2). 10 Related work: DALI Dali is a LP-like language for programming agents. E:> Cond, A1, A2,…,An. A1:> Cond, A3, A4, ... A1:> A5, A6,…. It is possible to define complex actions. E is always a basic event, no event algebra. It is not possible to update rules. Unlike ERA, DALI is implemented. 11 Future works Transaction ERA. To provide the possibility to execute actions as transactions. Local transactions are ACID. Transactions involving external iterations relays on compensation. Possibility to use ACID sub-transactions and compensation in the same transaction. 12 Thanks! 13 Future works A dynamic rule is of the form effect(F ← Body) ← Cond. It means: if Cond is true in this state, then F ← Body is a rule in the next state (and only in that one). Cond are conjunctions of fluent or action literals. effect(F ← Body) ← Cond is a macro for: F ← Body, event(F ← Body). assert(event(F ← Cond). ) ← Cond. assert( not event(F ← Cond). ) ← event(F ← Cond) , not assert(event(F ← Cond)). An inertial declaration is of the form inertial(I) ,where I is a set of fluents which means: all the fluents in I are inertial, i.e. their truth value is preserved from one state to the next one unless changed by the effect of an action. Inertial(I) is a macro for: assert(prev(F)) ← F. assert(not prev(F)) ← not F. 14 Example: The Yale shooting problem There is a single-shot gun which is initially unloaded, and some turkeys. I: We can load the gun and shoot to one of the turkeys. If we shoot, there is a bang, the gun becomes unloaded, we can hit the turkey or miss it, if we hit a turkey it dies. A turkey moves iff it is alive. initialize{loaded, moving(X),dead(X) missed(X), hit(X)} D: effect(loaded.) ← load. effect(not loaded.) ← shoot(X). effect (bang.) ← shoot(X). effect(hit(X) ← not missed(X) ) ← shoot(X). effect (missed(X) ← not hit(X) ) ← shoot(X). inertial (loaded) inertial (dead(X)) moving(X) ← not dead(X). dead(X) ←hit(X). 15 Comparisons with other works We compared EAPs with the existing action description languages A , B and the (definite fragment of) C . For every language we find a linear and modular translation into EAPs: Hence EAPs are at least as expressive as the considered languages. 16 Describing changes In practical situations, it may happen that the very rules of the domain change with time EAPs are of Evolp programs. So, they can be updated by external events. It is possible to update static rules and the descriptions of the effects of an action. We consider again the Yale shooting problem. Let us now consider that after some shots, we acquire rubber bullets. We can now either load the gun with normal bullets or with a rubber bullet, if we shoot with a rubber loaded gun, we never kill a turkey. E1: assert(initialize (rubber_loaded)) D1: effect (not dead(X) ← hit(X)) ← rubber _loaded. effect (not rubber _loaded) ← shoot(X). E2: assert(D1) Inertial(rubber_loaded) 17 Describing changes (cont.) We use a new macro construct not effect(r) to prevent the effects of an actions under some conditions Suppose the cylinder of the gun becomes dirty and, whenever one shoots, the gun may either work properly or fail. If it fails the shoot action has no real effect. E1: assert(initialize{ fails, work}) D2: effect (fails ← not work.) ← shoot(X). effect (work ← not fails.) ← shoot(X). effect (loaded ← fails.) ← loaded. effect(rubber_loaded ← fails) ← rubber_ loaded. effect(loaded←fails)← loaded. E2: assert(D2) not bang ← fails. not missed(X) ← fails. not hit(X) ← fails. 18 Conclusions Starting for the language for LP updates Evolp, we have defined the new action description language EAPs. We have shown how to use EAPs by an example. We have compared EAPs with existing work by showing translations of action languages A , B and C into EAPs. We have shown peculiar capabilities of EAPs to handle changes in the rules of the considered domain and provided an example of such usage. 19