individual and collective intention recognition combined
play

Individual and Collective Intention Recognition Combined with - PowerPoint PPT Presentation

Evolution Prospection Intention Recognition Home Ambient Intelligence Anytime Intention Recognition Individual and Collective Intention Recognition Combined with Evolution Prospection Lus Moniz Pereira and Han The Anh CENTRIAUNL


  1. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition 1 Evolution Prospection Evolution Prospection Agents Main Constructs 2 Intention Recognition Individual Intention Recognition Collective Intention Recognition 3 Home Ambient Intelligence Proactive Support Security and Emergency 4 Anytime Intention Recognition Model Construction Anytime Intention Recognition Algorithms logo 14 / 113 L.M. Pereira and T. A. Han Intention Recognition

  2. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Approach to Intention Recognition Intention recognition (IR): process by which an agent becomes aware of the intention of others. Main stream of IR: reducing to plan generation – generating conceivable plans achieving intentions and choosing ones matching observations. Main problem: finding initial set of possible intentions the planner has to work with. This set should depend on situation at hand: generating plans for all possible intentions is unrealistic. Accompanying paper EPIA-09: http: logo //centria.fct.unl.pt/~lmp/publications/online-papers/EPIA09-intention-recognition.pdf 15 / 113 L.M. Pereira and T. A. Han Intention Recognition

  3. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition New Approach We propose a new approach, using situation-sensitive Causal Bayes Nets – CBNs that change according to situation under consideration, itself subject to change, in order to: compute likelihood of intentions conditional on observations, and then filter out the much less likely ones. CBNs are modelled using P-log. A logical component is added to dynamically compute situation specific probabilistic information, to update into P-log program. logo 16 / 113 L.M. Pereira and T. A. Han Intention Recognition

  4. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Advantages of New Approach Plan generator only needs to deal with remaining relevant intentions. From likelihood value of intentions, the recognizing agent can see which are more likely and worth addressing first – especially important in case of quick decision making. Comparing to approaches using BNs solely, combining with plan generation guides recognition process: which actions should be checked for whether they were hiddenly executed, from their side-effects, say. logo 17 / 113 L.M. Pereira and T. A. Han Intention Recognition

  5. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Network structure Intentions are intermediate nodes whose ancestor nodes represent causes that give rise to them. When their prior probabilities can be specified without considering the causes, intentions are top nodes. Observed actions are children of the intentions that causally affect them. Observable effects are bottom nodes. They can be children of observed action nodes, of intention nodes, or of unobserved actions that might cause effects added as children of intention nodes. The causal relations, the conditional probability distribution (CPD) and the prior probabilities of top nodes are specified by domain logo experts. They may also be learnt automatically. 18 / 113 L.M. Pereira and T. A. Han Intention Recognition

  6. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition An Example: Fox-Crow Fable There is a crow, holding a cheese. A fox, being hungry, approaches the crow and praises her, hoping the crow will sing and the cheese will fall down near him. Unfortunately for the fox, the crow is very intelligent, having the ability of intention logo recognition. 19 / 113 L.M. Pereira and T. A. Han Intention Recognition

  7. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Fox’s Intentions CBN logo 20 / 113 L.M. Pereira and T. A. Han Intention Recognition

  8. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition P-log The computation in CBNs can be automated by using P-log P-log – a declarative language based on a logic formalism for probabilistic reasoning It uses Answer Set Programming as its logical and Causal Bayes Nets as its probabilistic foundations. logo 21 / 113 L.M. Pereira and T. A. Han Intention Recognition

  9. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Fox’s Intentions CBN in P-log Two sorts, bool and fox _ intentions , represent boolean values and set of Fox’s intentions bool = {t,f}. fox_intentions = {food,please,ter}. Attributes hungry _ fox , friendly _ fox , praised and i state that first 3 have no domain parameter and are boolean, and last maps each Fox intention to a boolean hungry_fox : bool. friendly_fox : bool. i : fox_intentions --> bool. praised : bool. These attributes are randomly distributed in their full ranges random(rh,hungry_fox,full). random(rf,friendly_fox,full). random(ri,i(I),full). random(rp,praised,full). logo 22 / 113 L.M. Pereira and T. A. Han Intention Recognition

  10. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Representing CPD in P-log: pa rule Probability distribution of top nodes hungry _ fox , friendly _ fox pa(rh,hungry_fox(t),d(1,2)). pa(rf,friendly_fox(t),d(1,100)). Probability distribution of Fox having intention food, conditional on its parents ( hungry _ fox , friendly _ fox ) pa(ri(food),i(food,t),d(8,10)) :- friendly_fox(t),hungry_fox(t). pa(ri(food),i(food,t),d(9,10)) :- friendly_fox(f),hungry_fox(t). pa(ri(food),i(food,t),d(1,100)) :- friendly_fox(t),hungry_fox(f). logo pa(ri(food),i(food,t),d(2,10)) :-friendly_fox(f),hungry_fox(f). 23 / 113 L.M. Pereira and T. A. Han Intention Recognition

  11. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Probability distribution of observation Praised , conditional on its parents ( i ( food ) , i ( please ) , i ( ter ) ) pa(rp, praised(t),d(95,100)) :- i(food, t), i(please, t). pa(rp, praised(t),d(6,10)) :- i(food, t), i(please, f). pa(rp, praised(t),d(8,10)) :- i(food, f), i(please, t). pa(rp, praised(t),d(1,100)) :- i(food, f),i(please,f),i(ter,t). pa(rp, praised(t),d(1,1000)) :- i(food,f),i(please,f),i(ter,f). logo 24 / 113 L.M. Pereira and T. A. Han Intention Recognition

  12. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Likelihood of intentions Probabilities of Fox having intention Food, Territory, Please given the observation that Fox praised Crow can be found in P-log with queries ? − pr ( i ( food , t ) | obs ( praised ( t )) , V 1 ) . V 1 = 0 . 9317. ? − pr ( i ( ter , t ) | obs ( praised ( t )) , V 2 ) . V 2 = 0 . 8836. ? − pr ( i ( please , t ) | obs ( praised ( t )) , V 3 ) . V 3 = 0 . 0900. logo 25 / 113 L.M. Pereira and T. A. Han Intention Recognition

  13. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Situation-sensitive CBNs CBNs should be situation-sensitive since using a general CBN for all specific situations of a problem domain is unrealistic and most likely imprecise. We provide a way to construct these CBNs use Logic Programming techniques, such as abduction, preferences, inductive learning, etc. to compute situation specific probabilistic information, and then update the specific information into a CBN general for the problem domain. logo 26 / 113 L.M. Pereira and T. A. Han Intention Recognition

  14. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Example (cont.) Logical component of Fox’s intentions CBN pa_rule(pa(ri(ter),i(ter,t),d(0,1)),[friendly_fox(t)]) :- territory(tree). pa_rule(pa(ri(ter),i(ter,t),d(1,100)),[friendly_fox(f)]) :- territory(tree). logo 27 / 113 L.M. Pereira and T. A. Han Intention Recognition

  15. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition If territory is tree, in the updated CBN the pa-rules for territory pa(ri(ter),i(ter,t),d(1,10)) :- friendly_fox(t). pa(ri(ter),i(ter,t),d(9,10)) :- friendly_fox(f). are replaced with pa(ri(ter),i(ter,t),d(0,1)) :- friendly_fox(t). pa(ri(ter),i(ter,t),d(1,100)) :- friendly_fox(f). The likelihood of the intentions Food, Territory, Please are: V 1 = 0 . 9407; V 2 = 0 . 0099; V 3 = 0 . 0908, respectively. Thus, most likely, the only surviving intention is food. logo 28 / 113 L.M. Pereira and T. A. Han Intention Recognition

  16. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Second Phase - Plan Generation The second phase of the intention recognition system is to generate conceivable plans that can achieve the most likely intentions surviving after the first phase. In our system we deploy ASCP – an ASP based conditional planner, re-implemented in XSB Prolog. logo 29 / 113 L.M. Pereira and T. A. Han Intention Recognition

  17. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition ASCP Syntax ASCP uses A c K – action language extending language A . Alphabet of A c K consists of a set of actions A and a set of fluents F . A literal is a fluent f ∈ F or its negation ¬ f . A fluent formula ϕ is a propositional formula constructed from the set of literals using operators ∧ , ∨ and/or ¬ . logo 30 / 113 L.M. Pereira and T. A. Han Intention Recognition

  18. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition To describe an action theory, 5 kinds of propositions used: 1 initially ( l ) : l holds in the initial situation. Initial situation is described by a set of these propositions. 2 executable ( a , ψ ) : a is executable in any situation in which ψ holds 3 causes ( a , l , φ ) : performing a in a situation in which φ holds causes l to hold in the successor situation 4 if ( l , ϕ ) : l holds in any situation in which ϕ holds. 5 determines ( a , θ ) : values of literals in θ will be known after a is executed. logo 31 / 113 L.M. Pereira and T. A. Han Intention Recognition

  19. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Example: Fox’s Plans for Food Initial situation initially(holds(crow,cheese)). initially(hungry(fox)). Executability conditions executable(eat(A,E),[holds(A,E)]) :- animal(A),edible(E). executable(sing(B),[accepted(B)]) :- bird(B). executable(praise(fox,A), []) :- animal(A). executable(grab(A,O),[holds(no,O)]) :- animal(A),obj(O). Causal laws causes(sing(B),holds(no,O),[holds(B,O)]):- bird(B),obj(O). causes(eat(A,E),neg(hungry(A)),[hungry(A)]) :- animal(A),edible(E). causes(grab(A,O),holds(A,O),[]) :- animal(A),obj(O). Knowledge propositions determines(praise(fox,B),[accepted(B),declined(B)]):-bird(B). logo Goal: goal([neg(hungry(fox))]) . 32 / 113 L.M. Pereira and T. A. Han Intention Recognition

  20. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Fox’s Plans to get food [ praise ( fox , crow ) , cases ( { accepted ( crow ) → [ sing ( crow ) , grab ( fox , cheese ) , eat ( fox , cheese )]; declined ( crow ) → ⊥} )] where ⊥ means no appropriate plans i.e, first, Fox praises Crow. If Crow accepts to sing, Fox grabs the dropped cheese and eats it. Otherwise, i.e. if Crow declines to sing, nothing happens. logo 33 / 113 L.M. Pereira and T. A. Han Intention Recognition

  21. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Elder Care domain In order to provide contextually appropriate help for elders, the assisting system needs to be able to observe elders’ actions, recognize their intentions, then provide suggestions achieving the recognized intentions We focus on dealing with the last two steps. Elders’ intention recognition is employed using the previously described system. Accompanying papers Springer Book Chapter 2010: http://centria.fct.unl.pt/~lmp/publications/online-papers/IR-EPA-CBNs-book.pdf INAP-09: http://centria.fct.unl.pt/~lmp/publications/online-papers/INAP09_Elder_Care.pdf logo 34 / 113 L.M. Pereira and T. A. Han Intention Recognition

  22. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition For providing suggestions to achieve recognized intentions: use our Evolution Prospection Agent (EPA) Prospectively look ahead into the future to choose the best course of evolution achieving the recognized intention, while aware of external environment and elder’s preferences Expectation rules and a priori preferences take into account physical health reports to guarantee only contextually safe healthy choices are generated; then, information like elder’s pleasure, interests, etc. are met by post preferences. Advance in handling and easiness of expressing preferences in EPA enable to closely take into account the elders’ preferences – to increase degree of acceptance of elders w.r.t. technological help – an important issue of Elder Care domain. logo 35 / 113 L.M. Pereira and T. A. Han Intention Recognition

  23. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition An Example An elder stays alone in his apartment. The intention recognition system observes that he is looking for something in the living room. In order to assist him, the system needs to figure out what he intends to find. The possible things are: something to read ( Book ); something to drink ( Drink ); TV remote control ( Rem ); the light switch ( Switch ). The states of the light and TV are observed. logo 36 / 113 L.M. Pereira and T. A. Han Intention Recognition

  24. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition logo 37 / 113 L.M. Pereira and T. A. Han Intention Recognition

  25. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Elder’s Intentions CBN in P-log Two sorts, bool and elder _ intentions , represent boolean values and set of the Elder’s intentions bool = {t,f}. elder_intentions = {book,drink,rem,switch}. Attributes thsty (thirsty), lr (like reading), lw (like watching), tv (tv on), lt (light on) have no domain parameter and get boolean values thsty:bool. lr:bool. lw:bool. tv:bool. lt:bool. Attribute i maps each Elder’s intention to a boolean value i:elder_intentions --> bool. These attributes are randomly distributed in their full ranges random(rth,thsty,full). random(rlr,lr,full). random(rlw,lw,full). random(rtv,tv,full). logo random(rl, lt, full). random(ri, i(I), full). 38 / 113 L.M. Pereira and T. A. Han Intention Recognition

  26. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Representing CPD in P-log: pa rule Probability distribution of top nodes pa(rth,thsty(t),d(1,2)). pa(rlr,lr(t),d(8,10)). pa(rlw,lw(t),d(7,10)). pa(rtv,tv(t),d(1,2)). pa(rl,lt(t),d(1,2)). Probability distribution of the Elder having intention book , conditional on its parents (all top nodes) pa(ri(b),i(b,t),d(0,1)) :-lt(f). pa(ri(b),i(b,t),d(0,1)) :-lt(t),tv(t). pa(ri(b),i(b,t),d(3,5)) :-lt(t),tv(f),lr(t),lw(t),thsty(t). pa(ri(b),i(b,t),d(13,20)):-lt(t),tv(f),lr(t),lw(t),thsty(f). pa(ri(b),i(b,t),d(7,10)) :-lt(t),tv(f),lr(t),lw(f),thsty(t). pa(ri(b),i(b,t),d(4,5)) :-lt(t),tv(f),lr(t),lw(f),thsty(f). pa(ri(b),i(b,t),d(1,10)) :-lt(t),tv(f),lr(f),lw(t). pa(ri(b),i(b,t),d(4,10)) :-lt(t),tv(f),lr(f),lw(f). logo Similarly for other intentions. 39 / 113 L.M. Pereira and T. A. Han Intention Recognition

  27. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Probability distribution of observation look , conditional on its parents ( i ( b ) , i ( dr ) , i ( rem ) , i ( sw ) ) pa(rla,look(t),d(99,100)):-i(b,t),i(dr,t),i(rem,t). pa(rla,look(t),d(7,10)) :-i(b,t) i(dr,t),i(rem,f). pa(rla,look(t),d(9,10)) :-i(b,t),i(dr,f),i(rem,t). pa(rla,look(t),d(6,10)) :-i(b,t),i(dr,f),i(rem,f). pa(rla,look(t),d(6,10)) :-i(b,f),i(dr,t),i(rem,t). pa(rla,look(t),d(3,10)) :-i(b,f),i(dr,t), i(rem,f). pa(rla,look(t),d(4,10)) :-i(b,f),i(dr,f),i(rem,t). pa(rla,look(t),d(1,10)) :-i(b,f),i(dr,f),i(rem,f),i(sw,t). logo pa(rla,look(t),d(1,100)) :-i(b,f),i(dr,f),i(rem,f),i(sw,f). 40 / 113 L.M. Pereira and T. A. Han Intention Recognition

  28. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Likelihood of intentions Probabilities of the Elder having intention book, drink, remote control, switch given the observation that he is looking for something and the states of light and TV (on or off), can be found in P-log with queries ? − pr ( i ( b , t ) | ( obs ( tv ( S 1 )) & obs ( lt ( S 2 )) & obs ( look ( t ))) , V 1 ) . ? − pr ( i ( dr , t ) | ( obs ( tv ( S 1 )) & obs ( lt ( S 2 )) & obs ( look ( t ))) , V 2 ) . ? − pr ( i ( rem , t ) | ( obs ( tv ( S 1 )) & obs ( lt ( S 2 )) & obs ( look ( t ))) , V 3 ) . ? − pr ( i ( sw , t ) | ( obs ( tv ( S 1 ))& obs ( lt ( S 2 )) & obs ( look ( t ))) , V 4 ) . where S 1 , S 2 are boolean values ( t or f ) instantiated during execution, depending on the states of the light and TV. logo 41 / 113 L.M. Pereira and T. A. Han Intention Recognition

  29. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition If the light is off ( S 2 = f ), then V 1 = V 2 = V 3 = 0, V 4 = 1 . 0, regardless of the state of the TV. If the light is on and TV is off ( S 1 = t , S 2 = f ), then V 1 = 0 . 7521 , V 2 = 0 . 5465 , V 3 = 0 . 5036 , V 4 = 0 . 0101. If both light and TV are on ( S 1 = t , S 2 = t ), then V 1 = 0 , V 2 = 0 . 6263 , V 3 = 0 . 9279 , V 4 = 0 . 0102. Thus, if one observes the light is off, definitely the elder is looking for the light switch. Otherwise, if one observes the light is on, no matter TV is on or off, the first three intentions book, drink, remote control still under consideration in next phase. The intention of looking for the light switch is very unlikely compared with the others, thus being ruled out. When there is light, one goes directly to the light switch if the logo intention is to turn it off, without having to look for it. 42 / 113 L.M. Pereira and T. A. Han Intention Recognition

  30. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Situation-sensitive CBNs The CBN may vary depending on some observed factors, e.g. the time of day, the current temperature, etc. We design a logical component for the CBN to deal with those factors: pa_rule(pa(rlk,lr(t), d(0,1)),[]):-time(T), T>0, T<5,!. pa_rule(pa(rlk,lr(t), d(1,10)),[]):-time(T), T>=5, T<8,!. pa_rule(pa(rlw,lw(t), d(9,10)),[]):- time(T), schedule(T,football),!. pa_rule(pa(rlw,lw(t), d(1,10)),[]):-time(T), (T>23; T<5),!. pa_rule(pa(rth,thsty(t), d(7,10)),[]):-temp(T), T>30,!. pa_rule(pa(rlk,lr(t), d(1,10)),[]):-temp(TM), TM >30,!. pa_rule(pa(rlw,lw(t), d(3,10)),[]):-temp(TM), TM>30,!. If body of rule true, corresponding P-log pa-rule is formed and updated into the CBN. logo 43 / 113 L.M. Pereira and T. A. Han Intention Recognition

  31. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Collective Intention Collective intention of a group of agents is not mere summation of individual intentions (Philosophy, AI): It involves a sense of acting and willing something together, cooperatively. It requires some “glue" supplement amongst the agents’ activity. “Glue" amongst agents, e.g. mutual beliefs and reactions: Agents hold reciprocal expectations. Thus, expectation reciprocal actions should be observed, and reactions to any lack thereof. Accompanying paper AAAI-FS-PAA-10: logo http://centria.fct.unl.pt/~lmp/publications/online-papers/PAA10_collective.pdf 44 / 113 L.M. Pereira and T. A. Han Intention Recognition

  32. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Bottom-up Approach Tuomela: collective intention = individual intentions + mutual beliefs. Briefly, agents A and B intend to do some task X cooperatively if these “glue" conditions for A—symmetrically for B—hold: (a) A intends to do his part of X (b) A believes B will do his part of X (c) A believes B believes A will do his part of X Kanno et al.: bottom-up approach to collective IR: constituents’ individual intentions and beliefs are inferred first; then, collective intention is inferred by checking for consistency amongst the inferred mental components. logo 45 / 113 L.M. Pereira and T. A. Han Intention Recognition

  33. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Bottom-up Method: Disadvantage Disadvantage: combinatorial explosion problem of possible combinations of individual intentions and beliefs for forming collective intentions. Given the situation at hand, each agent may have several conceivable intentions and beliefs, but not many of the combinations partake of conceivable collective intentions. logo 46 / 113 L.M. Pereira and T. A. Han Intention Recognition

  34. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Collective Intention Recognition We propose a top-down method for collective intention recognition, based on Searle’s account Searle’s Account Collective intention is non-summative, but remains individualistic. Having the presuppositions of mutual awarenesses or beliefs, posit the existence of a virtual plural agent that holds the collective intention This virtual plural agent is ascribed all the activities of the group of agents. logo 47 / 113 L.M. Pereira and T. A. Han Intention Recognition

  35. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Collective Intention Recognition Method Collective IR is thus reduced to individual IR plus checking if there are actions reflecting mutual expectations amongst agents: 1 First step: From the observations (actions or their effects in the environment of all agents in the group) infer intentions, as if these observations are of the plural agent; 2 Second step: Figure out which of the recognized intentions is a genuine collective intention, by checking for actions reflecting the mutual expectations between the agents, while ignoring any irrelevant actions for each considered intention. logo 48 / 113 L.M. Pereira and T. A. Han Intention Recognition

  36. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Example Example (Elder Care Couple) A couple of elderly people are alone in their apartment. The IR system observes they are both now in the kitchen. The man looking for something, the woman holding a kettle. In order to assist them, the system needs to figure out what they intend to do, cooperatively or separately. Possible collective intentions: “making a drink" or “cooking". The first step can be done by any existent individual IR method. We employ our previous system for this purpose, which relies on a Bayesian Network connecting possible logo intentions to actions and circumstances. 49 / 113 L.M. Pereira and T. A. Han Intention Recognition

  37. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition CBN: Couple’s Collective Intentions logo 50 / 113 L.M. Pereira and T. A. Han Intention Recognition

  38. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Computing Likelihood with P-log The probabilities of the elders having the collective intentions of cooking (cook) and making a drink (mD), given the man is looking for something and the woman holding a kettle, are computed with these P-log queries, respectively: ? − pr ( i ( cook , t ) | ( obs ( look ( t ) & obs ( holdKettle )) , V 1 ) . ? − pr ( i ( mD , t ) | ( obs ( look ( t ) & obs ( holdKettle )) , V 2 ) . Their results are: V 1 = 0 . 478 ; V 2 = 0 . 667. Meaning the collective intention of making a drink is more likely and should be examined first for confirmation. However, it is still necessary to look at the other collective intention as it is not much less likely. logo 51 / 113 L.M. Pereira and T. A. Han Intention Recognition

  39. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Situation-Sensitive CBN Example (Elder Care, cont’d) In this scenario, the Casual Bayesian Network (CBN) may vary depending on observed factors, e.g. time of day, of elders’ last drink or last meal. We design a logical component for the CBN to deal with these factors: pa_rule(pa(hg(t),d_(0,1)),[]):- time(T), eat(T1), T-T1<1. pa_rule(pa(hg(t),d_(9,10)),[]):- time(T), last_eating(T1),T-T1>3. pa_rule(pa(thty(t),d_(1,10)),[]):- time(T), drink(T1), T1-T<1. pa_rule(pa(thty(t),d_(9,10)),[]):- time(T), last_drink(T1),T1-T>3. logo 52 / 113 L.M. Pereira and T. A. Han Intention Recognition

  40. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Suppose current time is 18: time ( 18 ) last time elders ate was half an hour ago: last _ eating ( 17 . 5 ) they have not had any drink for 3 hours, e.g.: last _ drink ( 14 ) These three facts are asserted. Hence, these two pa _ rule / 2 literals become true, and are updated into the general CBN: pa_rule(pa(hg(t),d_(0,1)),[]). pa_rule(pa(thty(t),d_(9,10)),[]). Now the result is: V 1 = 0 . 0206 ; V 2 = 0 . 9993. This time around, only the collective intention of making a drink should be sought for confirmation in the second stage, as the one of cooking is quite unlikely. logo 53 / 113 L.M. Pereira and T. A. Han Intention Recognition

  41. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Second Step: Confirming Collective Intention This step aims to confirm whether the recognized intention is a genuine collective intention of the group. Let { a 1 , ...., a n } be the set of agents and A the plural agent standing for them (i.e. having all their presuppositions and actions). Suppose W is an intention of A , recognized in the first step. P = [ p 1 , ..., p k ] is a plan achieving W. We determine an assigned subplan to each agent towards achieving the collective intention of the whole group, by looking at each agent’s actions. logo 54 / 113 L.M. Pereira and T. A. Han Intention Recognition

  42. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Grouping Recall that P = [ p 1 , ..., p k ] is a plan for achieving W. Let s i be the first action of agent a i in P . Determine indices d i such that p d i = s i . Group the agents with same first action, i.e. with same index d i . They are doing the same task or at least some part of it together. Suppose we obtain m groups g 1 , ..., g m : g t is responsible for the subplan [ p j t + 1 , ..., p j t + 1 ] where 1 ≤ t ≤ m and 0 = j 1 < j 2 < ... < j m + 1 = k . Grouping is unique for a given set of agents and a given plan. logo 55 / 113 L.M. Pereira and T. A. Han Intention Recognition

  43. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Checking Expectation Actions To check if there are expectation actions reflecting agents’ mutual expectations, consider 2 cases: mutual expectations between agents in a group, and between agents in consecutive groups The number of interactions amongst the agents that need to be observed (e.g. by an activity recognition system) is considerably reduced. The number of possible expectation actions between two particular agents is reduced and becomes better specified. One group may interact (produces or consumes results) with more than another group, but that is not considered here. logo 56 / 113 L.M. Pereira and T. A. Han Intention Recognition

  44. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Inside Group When some agents are assigned or intended to do something together, they expect from each other to do some task. “complain" is an expectation action when a group member “deviates" from the task without an “inform" action. Example : two people intend to walk together, but if one changes direction without informing the other, he will be subject to complaint. This distinguishes it from the situation of walking in step by coincidence.. When agents work cooperatively, we usually observe a “help" action. This type of action is often preceded by a “complain" action. The term “deviation" should be specified for concrete application domains. E.g., in the spatial domain, an agent is said to deviate if he is not in his assigned position or does not logo move in the right direction, or does so at wrong speed. 57 / 113 L.M. Pereira and T. A. Han Intention Recognition

  45. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Between Consecutive Groups When agents are in consecutive groups: agents from the second would expect some result from the first, who in turn expect the agents from the second to use their result. If the agents from one of the groups “deviated" from their task or did not finish it as assigned, the agents from the other group would “complain". Let result t be the assigned result of group g t . Usually, this result comes from the last action, p j t + 1 , of the group subplan. Assume a given set of possible actions reflecting expectation of result t of action p j t + 1 , denoted by expect _ result t , and a given set of possible actions reflecting expectation of using result t , denoted by expect _ use t . logo 58 / 113 L.M. Pereira and T. A. Han Intention Recognition

  46. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Then, we say these two groups of agents are working towards achieving a collective intention either if agents in g t + 1 have some action belonging to expect _ result t , or agents in g t have some action belonging to expect _ use t , or there occur “complain" actions from one of the groups. As long as the collective intentionality inside each group is confirmed, a single expectation action observed between the two groups is enough to conclude that they have a collective intention. Usually, it is useful to identify one of the result-delivery and result-receiver agents. logo 59 / 113 L.M. Pereira and T. A. Han Intention Recognition

  47. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Advantages of Grouping Method: 1 Reduces number of interactions needing to be observed Set of n agents: number of interactions to be observed n ( n − 1 ) . 2 Grouping method: m groups of n j ( 1 ≤ n j ≤ n , 1 ≤ j ≤ m ) agents, where � m j = 1 n j = n . The number of interactions to be observed m n j ( n j − 1 ) ) + m − 1 = n ( n − 1 ) � ( − S 2 2 j = 1 where S = ( � 1 ≤ j < k ≤ m n j n k ) − m + 1 is the number of interactions not needing observation, compared with the case without grouping. If groups are equally divided, we reduce approximately m times logo the number of interactions to be observed. 60 / 113 L.M. Pereira and T. A. Han Intention Recognition

  48. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Advantages of Grouping Method: 2 Focuses on smaller groups, with smaller sets of possible expectation actions Divides big set of agents into smaller groups: more easily observed, e.g. by an activity recognition system. Specifies which are the expectation actions requiring recognition between a particular pair of agents: Are they in the same group? Are they in consecutive groups? Or else? Without grouping, a bigger set of possible expectation actions needs to be considered for any pair. logo 61 / 113 L.M. Pereira and T. A. Han Intention Recognition

  49. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Elder Care: General Case In order to provide appropriate assistance: When observing only individual activity (e.g. when the other is absent from the apartment): recognize individual intentions. When observing both elders’ activity: recognize whether there is collective intentionality. If no collective intentionality is detected, the system then should then perform individual IR. logo 62 / 113 L.M. Pereira and T. A. Han Intention Recognition

  50. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition An Example: Couple of Elder Example (Confirming Collective Intention) Suppose “making a drink" was the collective intention recognized in the first step. We check if it is a genuine one. Plan achieving the intention: [take kettle, fill up with water, boil water, look for tea or coffee, put into water] woman’s subplan: [take kettle, fill up with water, boil water] man’s subplan: [look for tea or coffee, put into water] . Woman’s assigned result: provide boiled water . Man’s expectation: boiled water . He may have “expect result" actions, e.g. ask whether the water is ready or get the water from the woman. Or, if after a while the woman had not boiled logo water or was doing something else, the man would complain. 63 / 113 L.M. Pereira and T. A. Han Intention Recognition

  51. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition If such “expect result" or “‘complain" action occurs, we conclude that they really have a collective intention of making a drink. Otherwise, e.g. the man does not show any expectation for water the woman has actually boiled, then we conclude that there was not a genuine collective intention, even if later he might use the boiled water for his own purpose. We emphasize that there are necessary actions showing mutual expectations of results and usage of results when agents have indeed a collective intention towards achieving some task. logo 64 / 113 L.M. Pereira and T. A. Han Intention Recognition

  52. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Now suppose the system has found there is no collective intention amongst the elders, and the man keeps looking for something. To assist him, the system should then figure out what is his individual intention. Example (Man’s intentions) Suppose the possibilities are: book, drink, remote control, cook, make drink. We have shown elsewhere how to proceed with individual intention recognition in our setting. logo 65 / 113 L.M. Pereira and T. A. Han Intention Recognition

  53. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Related Work We are not the first to suggest the use of the plural agent concept: most works in multi-agent plan recognition rely on the assumption that the plan is carried out by a single entity (team, group, troop). However, none of them has addressed the necessary cognitive underpinnings amongst the constituents (mutual beliefs or mutual awarenesses) in order to confirm the existence of a collective intention. That is however the main concern of the philosophical community regarding collective intentionality. logo 66 / 113 L.M. Pereira and T. A. Han Intention Recognition

  54. Evolution Prospection Intention Recognition Individual Intention Recognition Home Ambient Intelligence Collective Intention Recognition Anytime Intention Recognition Consequently, extant work is restricted to considering only sets of agents with an initially assigned collective intention (football team, army troop). In Elder Care domain concerning multiple users: elders’ actions may accidentally form a plausible plan for a conceivable intention, but each of them is following his/her own intention. Our method supplements existing ones that deal with arbitrary sets of agents; without an initially assigned collective task or intention. logo 67 / 113 L.M. Pereira and T. A. Han Intention Recognition

  55. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition 1 Evolution Prospection Evolution Prospection Agents Main Constructs 2 Intention Recognition Individual Intention Recognition Collective Intention Recognition 3 Home Ambient Intelligence Proactive Support Security and Emergency 4 Anytime Intention Recognition Model Construction Anytime Intention Recognition Algorithms logo 68 / 113 L.M. Pereira and T. A. Han Intention Recognition

  56. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Proactively Providing Support AmI key feature: system should take initiative to help. In order to provide contextually appropriate help, e.g. for elders, the assisting system needs being able to: observe the elders’ actions, 1 recognize his/her intentions, or their collective intention , 2 provide suggestions or help for achieving the recognized 3 intentions (Evolution Prospection Agents—EPA). Accompanying paper AITAmI-10: http://centria.fct.unl.pt/~lmp/publications/online-papers/AITAmI10.pdf logo 69 / 113 L.M. Pereira and T. A. Han Intention Recognition

  57. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Example (Elder Intentions) An elder stays alone in his apartment. One day, the Burglary Alarm is ringing. IR system observes that he is looking for something. In order to assist him, it needs to figure out what he intends to find. Possible things are: Alarm button ( AlarmB ) Contact Device ( ContDev ) Defensible Weapons ( Weapon ) Light switch ( Switch ) logo 70 / 113 L.M. Pereira and T. A. Han Intention Recognition

  58. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Bayesian Network for Intention Recognition logo 71 / 113 L.M. Pereira and T. A. Han Intention Recognition

  59. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Security for Home AmI 1 Security in terms of Burglary Alarm systems 2 Security in terms of health and well-being of residents logo 72 / 113 L.M. Pereira and T. A. Han Intention Recognition

  60. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Security: Burglary Alarm System Burglary Alarm technology has been based on sensing and recognizing very last actions of an intrusion plan (e.g. “breaking the door"). It may be too late to provide appropriate protection. Need to guess in advance possibility of intrusion from very first observed actions of potential intruders. This information can be sent to the carer to get prepared (e.g. turn on the light to scare off burglars). Two-stage Intention Recognition: From the observed actions the CBN computes likelihood of conceivable intentions If worrisome enough, carer should be informed of potential logo intrusion. 73 / 113 L.M. Pereira and T. A. Han Intention Recognition

  61. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Security: Health and Well-being AmI systems need to be able to prevent hazardous situations, which usually come from dangerous ideas or intentions (e.g. take bath when drunk, drink alcohol while not permitted, or even commit suicide) of the assisted persons, especially those with mental impairments. To this end, guessing their intentions from the very first relevant behaviors is indispensable for takeing timely actions. Our IR system: CBN computes how likely there is a dangerous intention, and carers should be informed in case it is likely enough, in order to get prepared. logo 74 / 113 L.M. Pereira and T. A. Han Intention Recognition

  62. Evolution Prospection Intention Recognition Proactive Support Home Ambient Intelligence Security and Emergency Anytime Intention Recognition Emergency Handling Wide range of emergency situations: recognizing intrusion intention; dangerous intentions of assisted persons; detecting fire; unconsciousness or unusualness in regular activities (e.g. sleep for too long), etc. Emergency handling in EPA Active goal rule for each emergency situation. For each goal: a list of possible actions, represented by abducibles, are available to form solutions. Users preferences: a priori prefs. for preferring amongst available actions; a posteriori prefs. for comparing solutions taking into account their consequences and utility; evolution-level preferences to compare further consequences. Expectation and counter expectations rules: encode pros and logo cons of users towards available actions. 75 / 113 L.M. Pereira and T. A. Han Intention Recognition

  63. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Outline 1 Evolution Prospection Evolution Prospection Agents Main Constructs 2 Intention Recognition Individual Intention Recognition Collective Intention Recognition 3 Home Ambient Intelligence Proactive Support Security and Emergency 4 Anytime Intention Recognition Model Construction Anytime Intention Recognition Algorithms logo 76 / 113 L.M. Pereira and T. A. Han Intention Recognition

  64. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Figure: On the left: Heinze’s tri-level decompositional model of intentional behavior of the intending agent: Intentional level; Activity level; and State level. On the right: intention recognition is the reversal of this process. logo 77 / 113 L.M. Pereira and T. A. Han Intention Recognition

  65. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Introduction We present an intention recognition (IR) model that is situation-dependent One aspect of intention is future-directedness: actions for the intention may be executed at a far distance in time. During that period the world may be changing, and an intention may be changed or abandoned. IR model is reconfigurable to take into account changes. incremental and anytime Bayesian Network (BN) is used as IR model. IR model is incrementally constructed as more actions are observed. An anytime BN inference algorithm is deployed to design an anytime IR algorithm. logo 78 / 113 L.M. Pereira and T. A. Han Intention Recognition

  66. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Summary We present an anytime algorithm for incremental intention recognition in a changing world. The algorithm is performed by dynamically constructing the intention recognition model on top of a prior domain knowledge base. The model is occasionally reconfigured by situating itself in the changing world and removing newly found out irrelevant intentions. Reconfigurable Bayesian networks are employed to produce the intention recognition model. Accompanying paper logo AAAI-FS-PAA-10: http://centria.fct.unl.pt/~lmp/publications/online-papers/PAA10_incremental.pdf 79 / 113 L.M. Pereira and T. A. Han Intention Recognition

  67. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Intention Recognition Model Construction The IR model is dynamically reconstructed as more actions are observed. The IR model is occasionally introspected (e.g. to remove found out irrelevant intentions) and reconfigured w.r.t. to the situation at hand. E.g., in the Elder Care domain, different elders have distinct conditions and habits to be taken into account to recognize their intentions. Also, place, time of day, temperature, etc. need to be considered. In the next slides we formalize the IR model based on the proposed structure, and provide some operators for incremental construction of the IR model. logo 80 / 113 L.M. Pereira and T. A. Han Intention Recognition

  68. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Definition (Bayes Networks) A Bayes Network is a pair a dag whose nodes represent variables and missing edges encode conditional independencies between the variables, and an associated probability distribution satisfying the Causal Markov Assumption (CMA), that is, variables are independent of their non-effects conditional on their direct own causes. Definition (Causal Bayes Network) A BN is causal if its associated probability distribution satisfies the condition specifying that if a node X of its dag is actively caused to be in a given state x (an operation written as do(x), e.g. in P-log syntax), then the probability density function changes to the one of the network obtained by cutting the links from X’s parents to X, logo and setting X to the caused value x. 81 / 113 L.M. Pereira and T. A. Han Intention Recognition

  69. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Bayesian Network for Intention Recognition Definition (Intention Recognition BN – IRBN) BN for intention recognition: triple W = �{ Cs , Is , As } , pa , P W � Cs, Is and As are sets of cause, intention and action nodes (binary random variables). pa maps a node to the set of its parent nodes: pa ( C ) = ∅ ∀ C ∈ Cs (causes are top node) pa ( I ) ⊆ Cs ∀ I ∈ Is (intention connect to its causes) pa ( A ) = Is ∀ A ∈ As (all intentions connect to each action). CPD tables are given by probability distribution P W P W ( X | pa ( X )) defines the probability of X conditional on pa ( X ) in W, for all X ∈ V W , where V W = Cs ∪ Is ∪ As . Furthermore, it is required that Cs = � I ∈ Is pa ( I ) , i.e. there is no logo isolated cause node in W. 82 / 113 L.M. Pereira and T. A. Han Intention Recognition

  70. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Unit Bayesian Net for Intention Recognition Our Intention Recognition method is performed by incrementally constructing an IRBN on top of a knowledge base of Unit IRBNs, defined as follows: Definition (Unit IRBN) The Bayesian Network for intention recognition for an action A, denoted by irBN ( A ) , is an IRBN where its set of intentions refers to a single action A. We denote by P A the probability distribution in irBN ( A ) . logo 83 / 113 L.M. Pereira and T. A. Han Intention Recognition

  71. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Knowledge Base We stipulate a reasonable set of assumptions for a domain knowledge base. Definition (Knowledge Base) The domain knowledge base KB consists of a set of actions AS and a set of unit IRBNs for every action in AS , satisfying that An intention I has the same set of parents (causes) and CPD table in all the unit IRBNs that it belongs to. C ( I ) denotes the set of parents of I. P KB ( I | C ( I )) defines CPD table of I. A cause C has the same prior probability distribution in all the unit IRBNs that it belongs to, denoted by P KB ( C ) . logo 84 / 113 L.M. Pereira and T. A. Han Intention Recognition

  72. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Operators for Handling IRBNs We present some operators for handling IRBNs and CPD tables Project of CPD table to a set of variables Combine IRBNs Remove a set of intentions from an IRBN Situate a IRBN in a situation logo 85 / 113 L.M. Pereira and T. A. Han Intention Recognition

  73. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Project of CPD Table This operator is useful later to reconfigure the model when new actions or intentions need to be removed from the model. Definition (Project of CPD Table) Let T be a CPD table defining P ( X | V ) : probability of random variable X conditional on set of random variables V. Project of T on strict subset V ′ of V ( V ′ ⊂ V ) according to assignment u to variables in U = V \ V ′ is a CPD table defining P ( X | V ′ , U = u ) (the part of T for V corresponding to U = u ) If variables of V are binary: proj ( T , V ′ ) denotes the project of T on V ′ according the assignment of all variables in U to false. logo 86 / 113 L.M. Pereira and T. A. Han Intention Recognition

  74. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Combination of Unit IRBNs Definition (Combination of Unit IRBNs) Let O = { A 1 , ..., A n } ⊆ AS ( n ≥ 0) be a set of actions, and irBN ( A i ) = �{ Cs i , Is i , { A i }} , pa i , P A i � be the unit IRBN for action A i (1 ≤ i ≤ n ). The IRBN of O is irBN ( O ) = �{ Cs O , Is O , O } , pa O , P O }� , where ∀ C ∈ Cs O , ∀ I ∈ Is O , 1 ≤ i ≤ n Is O = � n i = 1 Is i Cs O = � I ∈ Is O C ( I ) pa O ( C ) = ∅ pa O ( I ) = C ( I ) pa O ( A i ) = Is O P O ( C ) = P KB ( C ) P O ( I | pa O ( I )) = P KB ( I | C ( I )) P O ( A i | pa O ( A i )) is defined by proj ( T , Is O ) , where T is the CPD table for A i in irBN ( A i ) . logo 87 / 113 L.M. Pereira and T. A. Han Intention Recognition

  75. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Combination of IRBNs Definition (Combination of IRBNs) Let W 1 = �{ Cs 1 , Is 1 , As 1 } , pa 1 , P 1 � , W 2 = �{ Cs 2 , Is 2 , As 2 } , pa 2 , P 2 � be two IRBNs. The combination of W 1 and W 2 is comb ( W 1 , W 2 ) = �{ Cs , Is , As } , pa , P W � , where ∀ C ∈ Cs , ∀ I ∈ Is , ∀ A ∈ As As = As 1 ∪ As 2 Is = Is 1 ∩ Is 2 Cs = � I ∈ Is C ( I ) pa ( C ) = ∅ pa ( I ) = C ( I ) pa ( A ) = Is P W ( C ) = P KB ( C ) P W ( I | pa ( I )) = P KB ( I | C ( I )) P W ( A | pa ( A )) is defined by the CPD table proj ( T , Is ) , where T is the CPD table for A in irBN ( A ) . logo 88 / 113 L.M. Pereira and T. A. Han Intention Recognition

  76. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Remove Intentions from IRBN If some intentions are found out to be irrelevant, e.g. when their likelihoods are very small, they are removed from the model. Definition (Remove Intentions from IRBN) Let W = �{ Cs , Is , As } , pa , P W � be an IRBN; R ⊂ Is . The result of removing the intentions in R from W is the IRBN remove ( W , R ) = �{ Cs R , Is R , As R } , pa R , P R � where ∀ C ∈ Cs R , ∀ I ∈ Is R , ∀ A ∈ As R As R = As Is R = Is \ R Cs R = � I ∈ Is R C ( I ) pa R ( C ) = ∅ pa R ( I ) = C ( I ) pa R ( A ) = Is R P R ( C ) = P KB ( C ) P R ( I | pa R ( I )) = P KB ( I | C ( I )) P R ( A | pa R ( A )) is defined by the CPD table proj ( T , Is R ) , where logo T is the CPD table for A in W. 89 / 113 L.M. Pereira and T. A. Han Intention Recognition

  77. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Situate an IRBN In a given situation, an IRBN is re-situated by recomputing the prior probabilities of the top nodes. Definition (Situate IRBN) Let W = �{ Cs , Is , As } , pa , P W � be an IRBN. We say W is situated into situation SIT if prior probabilities of top nodes of W, i.e. P W ( C ) ( C ∈ Cs ) , are recomputed according to SIT. The result is an IRBN: situate ( W , SIT ) = �{ Cs , Is , As } , pa , P S � P S ( C ) ( C ∈ Cs ) are the new prior probabilities of top nodes, resulting from re-computation according to SIT. P S ( X | pa ( X )) = P W ( X | pa ( X )) ∀ X ∈ Is ∪ As . logo 90 / 113 L.M. Pereira and T. A. Han Intention Recognition

  78. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Algorithm 1. KB: domain knowledge base; AS: set of actions in KB. Repeat these steps until one intention remains in the IRBN or time limit is reached; in the latter case, the most likely intention in the previous cycle is the final result. 1 Let O ⊆ AS : set current observed actions, W: current IRBN. Let W ′ = comb ( W , irBN ( O )) . If O is the set of initially observed actions: W ′ = irBN ( O ) . 2 If there is an (expert defined) “salient" intention in W ′ , situate it according to the situation at hand curSIT : situate ( W ′ , curSIT ) = W ′′ ; otherwise, the IRBN remains the same: W ′′ = W ′ . 3 Compute the likelihood of each intention in W ′′ , conditional on the set of current observed actions in W ′′ . Remove logo intentions much less likely than the others. 91 / 113 L.M. Pereira and T. A. Han Intention Recognition

  79. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Algorithm 1: Some Remarks If an observed action makes the set of conceivable intentions empty, the action is considered irrelevant and discarded. At any cycle, if the likelihood of all intentions are very small: the sought for intention is abandoned. The causes and actions do not support or force the intending agent to keep pursuing his/her intention anymore. logo 92 / 113 L.M. Pereira and T. A. Han Intention Recognition

  80. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Anytime Intention Recognition Definition (Anytime Algorithm) An algorithm is anytime if it can produce a solution in a given time T and the quality of solutions improves with time after T . In Alg. 1, the standard BN inference algorithm is applied. If the time limit is reached within the first cycle, it cannot provide any IR decision. Otherwise, using an anytime BN inference algorithm provides an anytime IR algorithm. logo 93 / 113 L.M. Pereira and T. A. Han Intention Recognition

  81. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Revising Knowledge Base Representation Currently, all conceivable intentions that may give rise to some action, regardless of the situation in which it is observed, figure in Unit IRBN for that action. This set is usually quite big. The set of conceivable intentions for an action should depend on the situation the action is observed. We propose two methods for revising the KB representation to tackle this problem Conditional Unit IRBNs. Situation-sensitive Intentions. logo 94 / 113 L.M. Pereira and T. A. Han Intention Recognition

  82. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Conditional Unit IRBNs We revise the current representation by providing a set of Unit IRBNs for each action, and a criterion for choosing the appropriate one for the situation at hand. Each Unit IRBN is accompanied by a precondition. The criterion must satisfy that in each situation only one Unit IRBN is chosen. Situation is encoded as a logic program. Difficulties: design a BN for each situation is costly; moreover, it is difficult to specify all situations from the beginning. logo 95 / 113 L.M. Pereira and T. A. Han Intention Recognition

  83. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Situation-sensitive Intentions Whether an intention may give rise to a particular action should be situation-dependent. Common sense reasoning can be employed for this purpose. We combine LP techniques for common sense reasoning. logo 96 / 113 L.M. Pereira and T. A. Han Intention Recognition

  84. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition KB is accompanied by a LP P KB to help decide which intentions are conceivable in a given situation. AS = { A 1 , ..., A N } : set of actions of KB. BNs = { W 1 , ..., W N } : set of unit IRBNs of KB, where W i = �{ Cs i , Is i , As i } , pa i , P i � ( 1 ≤ i ≤ N ) . Intention I is conceivable w.r.t. action A if it is expected in the given situation and there is no expectation to the contrary. Thus, for 1 ≤ i ≤ N and I ∈ Is i , P KB contains rule: conceivable ( I ) ← A i , expect ( I ) , not expect _ not ( I ) Furthermore, for each I ∈ � N i = 1 Is i , P KB contains two rules: expect ( I ) ← Cond 1 . expect _ not ( I ) ← Cond 2 . logo 97 / 113 L.M. Pereira and T. A. Han Intention Recognition

  85. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Suppose action A i (1 ≤ i ≤ N ) is observed. The current situation is encoded by logic program SIT . To compute set of conceivable intentions that may give rise to A i , we use XSB Prolog built-in findall/3 predicate to detect all true conceivable/1 atoms of program P KB ∪ SIT ∪ { A i ←} . Suppose O is the obtained set of conceivable intentions. The IRBN obtained by removing the other intentions from W i , i.e. remove ( W i , Is i \ O ) , is used for intention recognition. If A i is not given in SIT, but is allowed as an abducible, then an observation action can be triggered to discover whether hypothetical A i took place, and SIT is updated accordingly. If explicit negation is available in SIT, known absence of actions may be recorded and duly taken into account. logo 98 / 113 L.M. Pereira and T. A. Han Intention Recognition

  86. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition An Example Example (Elder Care) An elderly person stays alone in his apartment. An IR system is set up to support his activities in the living room. Once, at night, the system observes that the elder is looking around for something ( look ). KB of the system has a Unit IRBN for this action. For illustration, consider a small set of conceivable intentions: Is = { book , water , weapon , lightSwitch } . The accompanying logic program P KB contains the following rules, for each concrete action and conceivable I ∈ Is : conceivable ( I ) ← look , expect ( I ) , not expect _ not ( I ) . logo 99 / 113 L.M. Pereira and T. A. Han Intention Recognition

  87. Evolution Prospection Intention Recognition Model Construction Home Ambient Intelligence Anytime Intention Recognition Algorithms Anytime Intention Recognition Example (cont.) Suppose in P KB the expectation and counter-expectation rules for these intentions are 1. expect ( book ) . expect _ not ( book ) ← light _ off . expect _ not ( book ) ← burglar _ alarm _ ring . 2. expect ( water ) . expect _ not ( water ) ← light _ off . expect _ not ( water ) ← burglar _ alarm _ ring . 3. expect ( weapon ) ← burglar _ alarm _ ring . expect _ not ( weapon ) ← light _ off . expect _ not ( weapon ) ← no _ weapon _ availabe . 4. expect ( lightSwitch ) . expect _ not ( lightSwitch ) ← light _ on , tv _ on . logo 100 / 113 L.M. Pereira and T. A. Han Intention Recognition

Recommend


More recommend