1. Fariba Sadri ICCL 08 ALP 1 Abductive Logic Programming (ALP) and its Application in Agents and Multi-agent Systems Fariba Sadri Imperial College London ICCL Summer School Dresden August 2008

2. Fariba Sadri ICCL 08 ALP 2 Contents ALP recap ALP and agents Abductive planning via abductive event calculus (AEC) Dynamic planning with AEC Abductive proof procedures – IFF and CIFF Hierarchical Planning Reactivity Plan repair Dialogues and Negotiation

3. Fariba Sadri ICCL 08 ALP 3 Abductive Logic Programming Recap from Bob Kowalski’s Slides Abductive Logic Programs have three components: P is a normal logic program. A is a set of abducible predicates. IC, the set of integrity constraints. Often, ICs are expressed as conditionals: If A1 &...& An then B A1 &...& An  B or as denials: not (A1 &...& An & not B) A1 &...& An & not B  false Normally, P is not allowed to contain any clauses whose conclusion contains an abducible predicate. (This restriction can be made without loss of generality.)

4. Fariba Sadri ICCL 08 ALP 4 ALP Semantics Recap from Bob Kowalski’s Slides Semantics: Given an abductive logic program,, an abductive answer for a goal G is a set Δ of ground atoms in terms of the abducible predicates such that: G holds in P  Δ IC holds in P  Δ or P  Δ  IC is consistent.

5. Fariba Sadri ICCL 08 ALP 5 Abduction is normally used to explain observations Recap from Bob Kowalski’s Slides Program P:Grass is wet if it rained Grass is wet if the sprinkler was on The sun was shining Abducible predicates A: it rained, the sprinkler was on Integrity constraint: it rained & the sun was shining  false Observation:Grass is wet Two potential explanations: it rained, the sprinkler was on The only explanation that satisfies the integrity constraint is the sprinkler was on.

6. Fariba Sadri ICCL 08 ALP 6 ALP and Agents Some references: Two ALP-based agent models: R.A. Kowalski, F. Sadri, From logic programming towards multi-agent systems. In Annals of Mathematics and Artificial Intelligence Volume 25, pages 391- 419 (1999) Later developments of this model in later papers by Bob Kowalski A. Kakas. P. Mancarella, F. Sadri, K. Stathis, F. Toni. Computational logic foundations of KGP agents, Journal of Artificial Intelligence Research. To appear

7. Fariba Sadri ICCL 08 ALP 7 ALP and Agents Agents can be seen as ALPs Logic Programs represent beliefs (more elaborate beliefs than Agent0 or AgentSpeak) Abducibles represent observations and actions Integrity Constraints represent –Condition-action rules for reactivity –Plan repair rules –Communication policies –Obligations and prohibitions

8. Fariba Sadri ICCL 08 ALP 8 ALP and Agents Abductive Logic Programs used for Planning Reactivity Plan repair Negotiation

9. Fariba Sadri ICCL 08 ALP 9 ALP and Agents A small planning example P: have(X) if buy(X) have(X) if borrow(X) A:buy, borrow, register (actions) IC: buy(X) & no-money  false buy(tv)  register(tv) Goal: have(tv) Δ1:buy(tv) & register(tv)(Plan 1) Δ2:borrow(pc)(Plan 2) If P also includes no-money then The only solution (only plan) is Δ2: borrow(pc).

10. Fariba Sadri ICCL 08 ALP 10 ALP for Planning Abductive Event Calculus (AEC) Some References Original Event Calculus R.A. Kowalski and M.J. Sergot (1986) A logic-based calculus of events. New Generation Computing, 4(1), 67-95. Abductive Event calculus P. Mancarella, F. Sadri, G. Terreni, F. Toni (2004) Planning partially for situated agents. In Leite, Torroni (eds.), Computational Logic in Multi-agent Systems, CLIMA V, Lecture Notes in Computer Science, Springer, 230-248. Also papers by M. Shanahan.

11. Fariba Sadri ICCL 08 ALP 11 ALP for Planning Abductive Event Calculus (AEC) Domain Independent Rules holds_at(F,T 2 )  happens(A,T 1 ), T 1 <T 2, initiates(A, T 1, F), ¬clipped(T 1, F, T 2 ) holds_at(F,T)  initially(F), 0  T, ¬clipped(0,F,T) clipped(T 1,F,T 2 )  happens(A,T), terminates(A,T,F), T 1  T<T 2

12. Fariba Sadri ICCL 08 ALP 12 ALP for Planning Abductive Event Calculus (AEC) holds_at(¬F,T 2 )  happens(A,T 1 ), T 1 <T 2, terminates(A, T 1, F), ¬declipped(T 1, F, T 2 ) holds_at(¬F,T)  initially(¬F), 0  T, ¬declipped(0,F,T) declipped(T 1,F,T 2 )  happens(A,T), initiates(A,T,F), T 1  T<T 2

13. Fariba Sadri ICCL 08 ALP 13 AEC Domain dependent rules Example: initially(¬have(money)) initiates(buy(X), T, have(X) )  ¬X=money initiates(borrow(X), T, have(X)) terminates(buy(X), T, have(money) ) precondition(buy(X), have(money))

14. Fariba Sadri ICCL 08 ALP 14 AEC Domain dependent rules Another Example Actions can be communicative actions : tell(Ag1, Ag2, Content, D) initially(no-info(tr-ag, arrival(tr101)) initiates(tell(X, tr-ag, inform(Q,I), D),T,have-info(tr- ag,Q,I))  holds_at(trustworthy(X),T) terminates(tell(X, tr-ag, inform(Q,I), D),T,no-info(tr- ag,Q))  holds_at(trustworthy(X),T) precondition(tell(tr-ag,X, inform(Q,I), D), have-info(tr- ag,Q,I))

15. Fariba Sadri ICCL 08 ALP 15 AEC Abducible happens Domain Independent Integrity Constraints holds_at(F,T) & holds_at(  F,T)  false happens(A,T)& precondition(A,P)  holds_at(P,T) and any Domain Dependent Integrity Constraints For example: holds_at(open-shop, T)  T≥9 & T≤18 happens(tell(a, Ag, accept(R), D), T) & happens(tell(a, Ag, refuse(R), D), T)  false

16. Fariba Sadri ICCL 08 ALP 16 ALP with Constraints Notice that AEC has times and constraint predicates, T1<T2, T1  T2, T1=T2, etc. We can extend the notion of abductive answer to cater for these, in the spirit of Constraint Logic Programming (CLP). Structure R consisting of –a domain D( R ) and –a set of constraint predicates and –an assignment of relations on D( R ) for each such constraint predicate.

17. Fariba Sadri ICCL 08 ALP 17 ALP with Constraints Semantics Given an ALP with constraints,, an abductive answer for a goal G is a Δ=(D, C) such that D is in terms of the abducible predicates, and for all groundings  of the variables in G, D, C such that  satisfies C (according to R ) G holds in P  D  IC holds in P  D  or P  D   IC is consistent.

18. Fariba Sadri ICCL 08 ALP 18 Back to AEC Given AEC and a goal G holds_at(g 1, T 1 ) & holds_at(g 2, T 2 ) & … & holds_at(g n, T n ) an answer for G is a parially ordered plan for achieving G. That is an answer for G is a  = (As, TC) As is a set of happens atoms, and TC is a set of temporal constraints, and  is an abductive answer to G wrt AEC with constraints.

19. Fariba Sadri ICCL 08 ALP 19 Example Domain dependent part : initially(¬have(money)) initiates(buy(X), T, have(X) )  ¬X=money initiates(borrow(X), T, have(X)) terminates(buy(X), T, have(money) ) precondition(buy(X), have(money)) happens(buy(X), T)  T≥9 & T≤18

20. Fariba Sadri ICCL 08 ALP 20 Example cntd. Goal: holds_at(have(pc), T) & T<12 One answer (plan):  1 =(happens(borrow(pc), T1), T1<T<12)  2 =(happens(borrow(money),T1), happens(buy(pc), T2) T1<T2, T2  9, T2<T<12)

21. Fariba Sadri ICCL 08 ALP 21 Example cntd. Goal: holds_at(have(pc), T) & holds_at(have(tv), T) How many answers (plans) are there ???? Some answers are:  1 =(happens(borrow(pc),T1), happens(borrow(tv),T2), T1<T, T2<T)  2 =(happens(borrow(money),T1), happens(buy(pc), T2), happens(borrow(money),T3), happens(buy(tv), T4), T1<T2, T2  9, T2 ≤18, T2<T3, T3<T4, T4  9, T4 ≤18, T4<T)

22. Fariba Sadri ICCL 08 ALP 22 Agent Cycle Messages Goals + ALP Environment + Observations Effects of actions Observe Act Reason

23. Fariba Sadri ICCL 08 ALP 23 An ALP agent cycle with explicit time to cycle at time T, record any observations at time T, resume thinking, giving priority to forward reasoning with the new observations, evaluate to false any alternatives containing sub-goals that are not marked as observations but are atomic actions to be performed at an earlier time, select sub-goals, that are not marked as observations, from among those that are atomic actions to be performed at times consistent with the current time, attempt to perform the selected actions, record the success or failure of the performed actions and mark the records as observations, cycle at time T+ n, where n is small. Note selecting an action involves both selecting an alternative branch of the search space and prioritising conjoint subgoals.

24. Fariba Sadri ICCL 08 ALP 24 Agent Cycle Now consider using AEC in an agent cycle. We will have: Observations Plan execution Interleaved planning and plan execution So we have to extend the AEC theory for this dynamic setting.

25. Fariba Sadri ICCL 08 ALP 25 (Dynamic) AEC Bridge Rules Bridge rules for connecting the AEC theory to observations: holds_at(F,T 2 )  observed(F,T 1 ), T 1  T 2, ¬clipped(T 1,F,T 2 ) holds_at(¬F,T 2 )  observed(¬F,T 1 ), T 1  T 2, ¬declipped(T 1,F,T 2 ) happens(A,T)  executed(A,T) happens(A,T)  observed(_, A[T], _) happens(A, T)  assume_happens(A, T) clipped(T 1,F,T 2 )  observed(¬F,T), T 1  T<T 2 declipped(T 1,F,T 2 )  observed(F,T), T 1  T<T 2

26. Fariba Sadri ICCL 08 ALP 26 (Dynamic) AEC Abducible assume_happens As well as abducibles we can also have a set of observables. These are fluents (or fluent literals) that can only be proved by being observed. The agent cannot plan to achieve them. In the KGP architecture they are called sensing goals. E.g. shop is open it is raining.

27. Fariba Sadri ICCL 08 ALP 27 Observations In general observations can involve: Observable predicates Abducible predicates Defined predicates – in which case the observation may be explained through abductions.

28. Fariba Sadri ICCL 08 ALP 28 (Dynamic) AEC Modify the set of domain independent integrity constraints: holds_at(F,T) & holds_at(  F,T)  false assume_happens(A,T) & precondition(A,P)  holds_at(P,T) assume_happens(A,T) & ¬executed(A,T), time_now(T’)  T>T’

29. Fariba Sadri ICCL 08 ALP 29 Example Domain dependent part : initially(¬have(money)) initiates(buy(X), T, have(X) )  ¬X=money initiates(borrow(X), T, have(X)) terminates(buy(X), T, have(money) ) precondition(buy(X), have(money)) precondition(buy(X), open-shop) observable

30. Fariba Sadri ICCL 08 ALP 30 Example cntd. Goal: holds_at(have(pc), T) & T<12 The agent can consider two alternative plans, to borrow pc to buy pc. Then an observation ¬open-shop (maybe a notice saying the shop is closed all day or until further notice) makes the agent focus on the first plan.

31. Fariba Sadri ICCL 08 ALP 31 Proof Procedures for Abductive Logic Programs (with constraints) Some references: CIFF: Endriss U., Mancarella P., Sadri F., Terreni G., Toni F., Abductive logic programming with CIFF: system description. The CIFF proof procedure for abductive logic programming with constraints. Both in Proceedings of Jelia 2004.

32. Fariba Sadri ICCL 08 ALP 32 Proof Procedures Endriss U., Mancarella P., Sadri F., Terreni G., Toni F., The CIFF proof procedure for abductive logic programming with constraints: theory, implementation and experiments. Forthcoming. A-System: Kakas A., Van Nuffelen B., Denecker M., A- system: problem solving through abduction. In Proceedings of the 17 th Internationla Joint Conference on Artificial Intelligence, 2001, 591-596.

33. Fariba Sadri ICCL 08 ALP 33 Proof Procedure : CIFF Builds on earlier work by Fung and Kowalski: The IFF proof procedure for abductive logic programming. Journal of Logic Programming, 1997. CIFF adds constraint satisfaction and a few other features for efficiency and extended applicability. Here we review the IFF proof procedure.

34. Fariba Sadri ICCL 08 ALP 34 IFF Roughly speaking : Given ALP : IFF reasons forwards with IC and backwards with the selective completion of P (wrt non-abducibles).

35. Fariba Sadri ICCL 08 ALP 35 IFF – Selective Completion Example of selective completion wrt non-abducibles): P:have(X) if buy(X) have(X) if borrow(X) A:buy, borrow (actions) IC: buy(X) & no-money  false Selective completion of P is: have(X) iff buy(X) or borrow(X) no-money iff false i.e. the abducibles are open predicates, the rest are completed. We can also designate the observables as open predicates.

36. Fariba Sadri ICCL 08 ALP 36 IFF – Backward reasoning (Unfolding) Backward reasoning (Unfolding) uses iff- definitions to reduce atomic goals (and observations) to disjunctions of conjunctions of sub-goals. E.G. A goal have(pc) will be unfolded to buy(pc) or borrow(pc).

37. Fariba Sadri ICCL 08 ALP 37 IFF – Forward reasoning (Propagation) Forward reasoning (Propagation) tests and actively maintains consistency and the integrity constraints by matching a new observation or new atomic goal p with a condition of an implicational goal p & q  r to derive the new implicational goal q  r. q can be reduced to subgoals by backward reasoning (unfolding) or can be removed by forward reasoning (propagation). r is added as a new goal (after p & q has been removed). r can then trigger forward reasoning or can be reduced to sub-goals. r is (in general) a disjunction of conjunctions of sub-goals.

38. Fariba Sadri ICCL 08 ALP 38 IFF – Forward reasoning (Propagation) In the example propagation will give: [buy(pc) & (no-money  false)] or borrow(pc). Suppose now we observe no-money. Another propagation step will give [buy(pc) & false] or borrow(pc).

39. Fariba Sadri ICCL 08 ALP 39 IFF – Some Other Inference Rules Logical equivalences replace a subformula by another formula which is both logically equivalent and simpler. These include the following equivalences used as rewrite rules: G & true iff GG & false iff false D or true iff trueD or false iff D. [true  D] iff D[false  D] iff true Splitting uses distributivity to replace a formula of the form (D or D') & G by (D & G) or (D'& G) Factoring “unifies two atomic subgoals P(t) & P(s) replacing them by the equivalent formula [P(t) & s=t] or [P(t) & P(s) & s  t]

40. Fariba Sadri ICCL 08 ALP 40 IFF - Negation Two versions: Negation re-writing: ¬P is re-written as p  false. Combining negation re-writing and negation as failure: Some negative literals are marked These are evaluated using special inference rules that provide the effect of NAF.

41. Fariba Sadri ICCL 08 ALP 41 IFF - Search Space The search space is represented by a logical formula, e. g. [happens(borrow(money), T1) & happens(buy(pc), T2) & T1<T2 & happens(borrow(tv), T3) & precondition(buy(pc),P)  holds_at(P,T2)] or [happens(borrow(pc),T)] & happens(happens(borrow(tv),T’)] Each disjunct is analogous to an alternative branch of a prolog-like search space.

42. Fariba Sadri ICCL 08 ALP 42 Notice - 1 Propagation together with unfolding allows ECA or condition-action rule type of behaviour. E & C  A E & C  G trigger via propagation evaluate via propagation or unfolding fires the action or goal

43. Fariba Sadri ICCL 08 ALP 43 Notice - 2 Factoring allows repeating actions at a later stage if an earlier attempt has not been effective. Scenario: e-shopping – logged on with one credit card, choose item, then card fails because there are not enough funds. Given : happens(logon(Card), T1) & happens(logon(visa), 5) & Rest(Card) planned action recorded observation factoring obtains: [happens(logon(Card), T) & Card=visa &T=5 & Rest(Card)] or [happens(logon(Card), T) & happens(logon(visa), 5) & (T  5 or Card  visa) & Rest(Card)]

44. Fariba Sadri ICCL 08 ALP 44 Notice – 2 cntd. So if Rest(Card) fails with Card=visa &T=5 the 2 nd disjunct allows further attempts: –either to logon with a different card at a later time –or to logon with visa at a later time Notice also that any work done in Rest(Card) in the first attempt is saved and available in the 2 nd disjunct, for example choice of item to buy.

45. Fariba Sadri ICCL 08 ALP 45 Semantics Let D be a conjunction of definitions in iff-form Let G be a goal, i.e.conjunction of literals Let IC be a set on integrity constraints Let O be a conjunction of positive and negative observations, including actions successfully or unsuccessfully performed by the agent.  G &  O is an answer iff all the below hold:  G and  O are both conjunctions of formulae in the abducible and constraint (e.g.=) (or observable) predicates D &  G &  O |= G D &  O |= O D &  G &  O |= IC.

46. Fariba Sadri ICCL 08 ALP 46 AEC for Hierarchical Planning So far you have seen AEC formalised for planning from 1 st principles. This may not be ideal in an agent setting. AEC lends itself to formalising plan libraries, macro-actions and hierarchical and progressive planning. We will look at some of these through examples.

47. Fariba Sadri ICCL 08 ALP 47 AEC for Hierarchical Planning A Reference: M. Shanahan, An abductive event calculus planner. The Journal of Logic Programming, Vol 44, 2000, 207-239

48. Fariba Sadri ICCL 08 ALP 48 AEC for Hierarchical Planning Example (from Shanahan’s paper) Robot mail delivery domain Primitive actions initiates(pickup(P), T, got(P))  holds_at(in(R),T), holds_at(in(P,R),T), initiates(putdown(P),T, in(P,R))  holds_at(in(R),T), holds_at(got(P),T) initiates(gothrough(D), T, in(R1))  holds_at(in(R2),T), connects(D,R2,R1) terminates(gothrough(D), T, in(R))  holds_at(in(R),T)

49. Fariba Sadri ICCL 08 ALP 49 AEC for Hierarchical Planning Example cntd. Compound action happens(shiftPack(P,R1,R2,R3),T1,T6)  happens(goToRoom(R1,R2),T1,T2), ¬clipped(T2,in(R2),T3), ¬clipped(T1,in(P,R2),T3), happens(pickup(P),T3), happens(goToRoom(R2,R3),T4,T5), ¬clipped(T3,got(P),T6), ¬clipped(T5,in(R3),T6), happened(putdown(P),T6), T2<T3<T4, T5<T6

50. Fariba Sadri ICCL 08 ALP 50 AEC for Hierarchical Planning Example cntd. goToRoom(R1,R2) maintain effect pickup(P) maintain effect goToRoom(R2,R3) maintain effect putdown(P) shiftPack (P, R1, R2, R3)

51. Fariba Sadri ICCL 08 ALP 51 AEC for Hierarchical Planning Example cntd. initiates(shiftPack(P,R1,R2,R3),T, in(P,R3))  holds_at(in(R1), T), holds_at(in(P,R2),T) Verifiable: The effects of compound actions should follow from the effects of their sub- actions. This can be verified formally, or, in this case by inspection.

52. Fariba Sadri ICCL 08 ALP 52 AEC for Hierarchical Planning Example cntd. happens(goToRoom(R,R),T,T) happens(goToRoom(R1,R3),T1,T3)  connects(D,R1,R2), happens(gothrough(D),T1), ¬clipped(T1,in(R2),T2), T1<T2 happens(goToRoom(R2,R3),T2,T3) initiates(goToRoom(R1,R2),T,in(R2))  holds_at(in(R1),T) Conditional and recursive compound action

53. Fariba Sadri ICCL 08 ALP 53 AEC for Hierarchical Planning in agent model Hierarchical planning in the agent model allows : Building heuristics and expertise into planning Generating actions in progressive order – first action first (as opposed to regressive order – last action first). Progressive planning fits well within an agent cycle: A partial plan can be executed and give useful results. Observe effect of actions and the state of the environment to decide whether it is worth continuing with that plan.

54. Fariba Sadri ICCL 08 ALP 54 ALP for Reactivity To specify reactions to the changes/events in the environment, similar to condition-action rules To specify plan repair steps To specify interaction policies

55. Fariba Sadri ICCL 08 ALP 55 specify reactions to the environment ALP allows and extends the active behaviour provided by ECA/Condition- Action rules and gives it semantics.

56. Fariba Sadri ICCL 08 ALP 56 ALP for Reactivity Domain Dependent Integrity Constraints in the ALP can be specified and used to achieve reactive behaviour. Triggers, Other-Conditions  Reaction Conjunction of: Observationsholds_at(F,T) Executed actions happens(A,T) Planned actions temporal constraints

57. Fariba Sadri ICCL 08 ALP 57 Examples Informal syntax in room R at time T & alarm sounds in R at T  leave R at T+1 (Smart Home Ambient Intelligence AmI- See for example work by Augusto et al on agent-based ECA-based AmI) P leaves kitchen at T & gas is on at T & ¬ X enters kitchen by T+15  raise alarm at T+15

58. Fariba Sadri ICCL 08 ALP 58 Plan Repair Most agent models that have plan repair use (ad hoc) ECA/Condition-Action rules for this purpose (see, for example 2APL/3APL). A reference for 3APL: M. Dastani, B, van Riemsdijk, F. Dignum, J-J. Meyer, A programming language for cognitive agents goal directed 3APL. Proceedings of the First Workshop on Programming Multiagent Systems: Languages, frameworks, techniques, and tools (ProMAS03) to be held at AAMAS'03, Melbourne, July 2003.

59. Fariba Sadri ICCL 08 ALP 59 Plan Repair An example from 2APL: Goto(R2,R1);X <- not pos(R2) and pos(R3)| {goto(R3,R1);X} An example from 3APL: G(on(X,Y))<- B(tooheavy(X) and ¬heavy(Z)) | G(on(Z,Y)

60. Fariba Sadri ICCL 08 ALP 60 Plan Repair Using ALP we can Formalise and use such ad hoc (active) rules for plan repair – examples seen earlier Achieve the behaviour of such rules just by executing AEC – example on next slide Derive such specialised plan repair rules from the general AEC theory and then use them explicitly – example on slide after next Simulate TR-Program type of behaviour by manipulating the search space and search strategy – work in progress

61. Fariba Sadri ICCL 08 ALP 61 Plan Repair Achieve the behaviour of such rules just by executing AEC: Example: Goto(R2,R1);X <- not pos(R2) and pos(R3)| {goto(R3,R1);X} In AEC there will be a goal, e.g pos(r1) requiring an action goto(Y,r1) with a precondition pos(Y).

62. Fariba Sadri ICCL 08 ALP 62 Plan Repair pos(r1) goto(Y,r1) pos(Y), goto(Y,r1) Y=r2, ¬clipped(pos(r2)), goto(Y,r1) observe(pos(r3)) This entails clipped(pos(r2)) and pos(r3) Y=r3, ¬clipped(pos(r3)), goto(Y,r1)

63. Fariba Sadri ICCL 08 ALP 63 Plan Repair Deriving specialised plan repair rules from the general AEC theory E-shopping Example using simplified notation : logged on with one credit card, choose item, then card fails because there are not enough funds. We want to use another card, but first we have to logout the first card. Specialised plan repair rule logon(Card1,T1) & logged-on(Card2, T2) & T2<T1  logout(Card2, T3) & T2<T3<T1 Such a rule is derivable from AEC and domain-dependent rules precondition(logon(Card), ¬logged-on(Card’)) terminates(logout(Card), T, logged-on(Card))

64. Fariba Sadri ICCL 08 ALP 64 Plan Repair How? Briefly : Given happens(logon(Card1),T1) holds(¬logged-on(Card2), T1) using the precondition IC holds(logged-on(Card2),T1)  false using the F and ¬F cannot hold together IC happens(logon(Card2),T2) & T2<T1 & ¬clipped(T2, logged-on(Card2), T1)  false using definition of holds happens(logon(Card2),T2) & T2<T1  clipped(T2,logged-on(Card2),T1) using negation re-writing happens(logon(Card2),T2) & T2<T1  happens(logoout(Card2),T3) & T2<T3<T1 using definition of clipped and terminates So: happens(logon(Card1),T1) & happens(logon(Card2),T2) & T2<T1  happens(logoout(Card2),T3) & T2<T3<T1

65. Fariba Sadri ICCL 08 ALP 65 ALP for negotiation Some references: Sadri F., Toni F., Torroni P.: An abductive logic programming architecture for negotiating agents, Jelia 02. Minimally intrusive negotiating agents for resource sharing, IJCAI 03. Dialogues for negotation: agent varieties and dialogue sequences, ATAL 01. Endriss U., Maudet N., Sadri F., Toni F.: Protocol conformance for logic-based agents, IJCAI 03. F. Sadri: Multi-agent Cooperative Planning and Information Gathering, 11 th International Workshop CIA 2007 on Cooperative Information Agents, September 2007, LNAI series by Springer Verlag.

66. Fariba Sadri ICCL 08 ALP 66 ALP for negotiation Dialogues between agents as a means of interaction Often based on fixed protocols (rules of interaction) Negotiation is one form of dialogue Others include (Classification of dialogues [Walton & Krabbe, 1995]) –Persuation –Information seeking

67. Fariba Sadri ICCL 08 ALP 67 ALP for negotiation Negotiation: Negotiation is “the process by which a group of agents communicate with each other to try and come to a mutually acceptable agreement on some matter”. [Bussman & Muller 1992] One reason agents may negotiate is for resource sharing and allocation.

68. Fariba Sadri ICCL 08 ALP 68 ALP for negotiation General Idea 1 Agent1 has a plan requiring resources A,B. It has A,E, and is missing B. Agent2 has a plan requiring resources D,E. It has B,C,D, and is missing E. Can you give me B? Yes, if you give me E.

69. Fariba Sadri ICCL 08 ALP 69 ALP for negotiation General Idea 2 Can you give me B? No, why do you want B? I have a goal G and a plan bla bla and I need B for it. Well, You can solve G with plan bla’ bla’ which needs C but not B. I can give you C if you give me E.

70. Fariba Sadri ICCL 08 ALP 70 ALP for negotiation General Idea 3 Can you give me B from 9 to 5? No, I can give it to you at 9 but I need it back at 3.

71. Fariba Sadri ICCL 08 ALP 71 ALP for negotiation Communication language Communication language tell(Ag1,Ag2,Content,D) Content can be : request request(give(R)) request(give(R),(Ts,Te) accept refuse promise promise(R, (T1,T2), (T3,T4)) Changechange(promise(R, (T1,T2), (T3,T4))) Challengechallenge(request(give(R)))....

72. Fariba Sadri ICCL 08 ALP 72 ALP for negotiation Interaction Policies Expressed as integrity constraints of the form P i & C  P i+1 Dialogue move Intended meaning: If the agent receives a move p i and the conditions C are satisfied in its KB then it generates a move P i+1.

73. Fariba Sadri ICCL 08 ALP 73 ALP for negotiation Interaction Policies Examples observed(C, tell(C,a,request(R),D,T1),T) & holds_at(have(R),T1) & holds_at(¬need(R),T1)  happens(tell(a,C,accept(request(R)),D,T1),T2) & T+5>T2>T observed(C, tell(C,a,request(R),D,T1),T) & holds_at(need(R),T1)  happens(tell(a,C,refuse(request(R)),D,T1),T2) & T+5>T2>T observed(C, tell(C,a,request(R),D,T1),T) & holds_at(¬have(R),T1)  happens(tell(a,C,refuse(request(R)),D,T1),T2) & T+5>T2>T

74. Fariba Sadri ICCL 08 ALP 74 Interaction Protocols These policies conform to the following simple protocol requestaccept refuse

75. Fariba Sadri ICCL 08 ALP 75 Interaction Protocols request refuse accept promise change promise accept promise

76. Fariba Sadri ICCL 08 ALP 76 ALP for negotiation Considerations : Policies conforming to protocols Characterising weak/strong conformance Properties of conversations induced by policies, for example: –Termination of conversations –Success of policies in ensuring a resource sharing solution is found if one exists