Intention reconsideration in theory and practice

Transcription

1 Intention reconsidertion in theory nd prctice Simon Prsons nd Ol Pettersson nd lessndro Sffiotti nd Michel Wooldridge bstrct. utonomous gents operting in complex dynmic environments need the bility to integrte robust pln execution with higher level resoning. This pper describes work to combine low level nvigtion techniques drwn from mobile robotics with delibertion techniques drwn from intelligent gents. In prticulr, we discuss the combintion of nvigtion system bsed on fuzzy logic with delibertor bsed on the belief/desire/intention (DI) model. We discuss some of the subtleties involved in this integrtion, nd illustrte it with n exmple. 1 INTRODUCTION Milou the robot works in food fctory. He hs to regulrly go nd fetch two food smples (potto crisps) from two production lines in two different rooms, nd, nd tke them to n electronic tester in the qulity control lb. Milou must now pln his next delivery. He decides to get the smple from first, since room is closer thn. While going there, however, he finds the min door to tht room closed. Milou knows tht there is nother door tht he could use, but he considers the desirbility of doing so. The lterntive route to is hrd for Milou, since it goes through long nrrow corridor which is usully cluttered with boxes. esides, doors usully do not sty closed for long. Hence, Milou decides to go to first, nd come bck to lter on. He goes to room, picks up the potto crisps nd returns. The door to is still closed, nd this time Milou hs no other choice thn tking the difficult route. He does so, obtins the desired crisps, nd finlly goes to the lb nd completes his tsk. Performing the bove tsk requires the bility to nvigte robustly in rel-world, unsimplified environments. Milou must be ble to relibly find his wy, keep trck of his own position, void ny obstcles in the cluttered corridor, nd so on. However, this tsk lso requires some higher level cpbilities, like resoning bout lterntive wys to perform given tsk, nd reconsidering vilble options in the fce of new events. The development of intelligent mobile robots nd their deployment in rel-world environments will criticlly depend on our bility to integrte these two spects of the utonomous nvigtion problem. Tody s reserch on mobile robotics hs produced lrge number of techniques for robust nvigtion in rel environments in the presence of uncertinty, for exmple [1, 5, 11]. These techniques typiclly focus on the nvigtion problem, nd do not involve bstrct resoning processes of the type encountered in the bove scenrio. On the other hnd, reserch in intelligent gency hs resulted in number of powerful theories for resoning bout ctions nd plns. Deprtment of Computer Science, University of Liverpool, Chdwick uilding, Pech Street, L69 7ZF, Liverpool, United Kingdom. pplied utonomous Sensor Systems, Deprtment of Technology, Örebro University, S Örebro, Sweden. Indeed, much of the reserch ctivity from the intelligent gent community in the mid-to-lte 1980s ws focussed round the problem of designing gents tht could chieve n effective blnce between delibertion (the process of deciding wht to do) nd mens-ends resoning (the process of deciding how to do it) [3]. One prticulrly successful pproch tht emerged t this time ws the belief-desire-intention (DI) prdigm [3, 7, 13]. The development of the DI prdigm ws to gret extent driven by rtmn s theory of (humn) prcticl resoning [2], in which intentions ply centrl role. Put crudely, since n gent cnnot deliberte indefinitely bout wht courses of ction to pursue, the ide is it should eventully commit to chieving certin sttes of ffirs, nd then devote resources to chieving them. These chosen sttes of ffirs re intentions, nd once dopted, they ply centrl role in future prcticl resoning [2, 4]. mjor issue in the design of gents tht re bsed upon models of intention is tht of when to reconsider intentions. n gent cnnot simply mintin n intention, once dopted, without ever stopping to reconsider it. From time-to-time, it will be necessry to check, (for exmple), whether the intention hs been chieved, or whether it is believed to be no longer chievble [4]. In such situtions, it is necessry for n gent to deliberte over its intentions, nd, if necessry, to chnge focus by dropping existing intentions nd dopting new ones. Clerly, n gent s intention reconsidertion policy will ffect its performnce, nd the optiml policy for given gent will be hevily dependent upon its environment. There hs been certin mount of work on this problem in the re of intelligent gents, from both forml [17] nd n experimentl [9, 16] perspective. However, most of this work hs concentrted on gents in environments which re rther simple when compred to the environment Milou opertes in. Indeed, to our knowledge, there hs been no work which ttempts to investigte intention reconsidertion in environments which re both complex nd dynmic. Our reserch ims to ddress this deficit, identifying suitble mechnisms nd strtegies for intention reconsidertion which work well when combined with the kind of low-level control mechnisms required by gents which operte in complex dynmic environments. This pper describes one pproch which combines robust nvigtion system bsed on fuzzy logic [14, 15] nd DI system for hndling intentions. efore presenting the combintion, however, we discuss the problem of intention reconsidertion with respect to the forml model developed in [17]. 2 THE FORML MODEL Following [17], our gents hve two min dt structures: belief set nd n intention set. n gent s beliefs represent informtion tht the gent hs bout its environment. Let be the set of ll beliefs. For the most prt, the contents of will not be of concern to us here. How-

2 ever, it is often useful to suppose tht contins formule of some logic, so tht, for exmple, it is possible to determine whether two beliefs re mutully consistent or not. n gent s ctions t ny given moment re guided by its intention set, nd its intentions my be thought of s sttes of ffirs tht the gent hs committed to bringing bout. These my be structured in some wy for instnce in hierrchy with high level intentions defined s set of lower level intentions nd my be ordered. Formlly, let I be the set of ll intentions. gin, we re not concerned here with the contents of I. s with beliefs, however, it is often useful to ssume tht intentions re expressed in some sort of logicl lnguge. n gent s locl stte will then be pir b i, where b is set of beliefs, nd i I is set of intentions. Let L I be the set of ll internl sttes of the gent. We use l (with nnottions: l l ) to stnd for members of L. If l b i, then we denote the belief component of l by b l, nd the intention component by i l. For the forml model we ssume fixed set of intentions which hve been generted from some set of desires in the usul wy [3]. gents do not operte in isoltion: they re situted in environments; we cn think of n gent s environment s being everything externl to the gent. We ssume tht the environment externl to the gent my be in ny of set E e e of sttes. For now we ssume tht n gent knows wht stte the environment is in, cknowledging tht, in future work, we will hve to tke ccount of the fct tht ny gent only hs prtil knowledge of the environment. Together, n gent nd its environment mke up system. The globl stte of system t ny time is thus pir contining the stte of the gent nd the stte of the environment. Formlly, let G E L be the set of ll such globl sttes. We use g (with nnottions: g g ) to stnd for members of G. Our gents hve four min functionl components, which together generte their behviour: next-stte function, met-level control function, delibertion function, nd n ction function. The next stte function cn be thought of s belief revision function. On the bsis of the gent s current stte nd the stte of the environment, it determines new set of beliefs for the gent, which will include ny new informtion tht the gent hs perceived. n gent s next-stte function thus relises whtever perception the gent is cpble of. Formlly, next-stte function is mpping E. The next component in our gent rchitecture is met-level control. The ide here is tht t ny given instnt, n gent hs two choices vilble to it. It cn either deliberte (tht is, it cn expend computtionl resources deciding whether to chnge its focus), or else it cn ct (tht is, it cn expend resources ttempting to ctully chieve its current intentions). Note tht we ssume the only wy n gent cn modify its intentions is through explicit delibertion. To represent the choices vilble to n gent, we will ssume set C d, where d denotes delibertion, nd denotes ction. The purpose of n gent s met-level control function it to choose between delibertion nd ction. If it chooses to deliberte, then the gent subsequently delibertes; if its chooses to ct, then the gent subsequently cts. Formlly, we cn represent such strtegies s functions L C. The delibertion process of n gent is represented by function tht, on the bsis of n gent s internl stte, determines new set of intentions. Formlly, we cn represent this delibertive process vi function! L " I. If n gent decides to ct, rther thn deliberte, then it is cting to chieve its intentions. To do so, it must decide which ction to perform. The ction selection component of n gent is essentilly function tht, on the bsis of the gent s current stte, returns n ction, which represents tht which the gent hs chosen to perform. Let c #$% &$'( be the set of ctions. Formlly, n ction selection function is mpping ) L c. Finlly, we define n gent to be 5-tuple * +! &), (- l./, where is met-level control function,! is delibertion function, ) is n ction selection function, is next-stte function, nd l.10 L is n initil stte. 3 MILOU IN THEORY Our intention in introducing this forml model is to shed light upon the problems one fces when ttempting to integrte high-level gent rchitecture like the DI model with the concrete requirements of mobile robot. Consider Milou once gin. In the bstrct terms used by the DI model we cn consider Milou to hve set of possible intentions: i c 2&34+2'56&784:9;4 i 2=<?>@32&;3C4:;DE6:2F6&DHGI2&3C2&D i f JK32&5LM56&784:9;4 i N 2=<?>@32&;3COPDEQ;R16&DEGI2&3C2&D i t 2=<?4+2&3%56&784:9;4 i b REDS2&D i l REDT2&DU2&;3VOW<?X These intentions re hierchiclly structured nd ordered, with i c being composed of i f followed by i t, nd i f being composed of i or i N long with i b nd followed by i l. The lowest level of these intentions hve corresponding ctions denoted by $ :$ N :$ b :$ l. Milou lso hs set of possible beliefs: b 4:;D?6:2'6&DHGY2&3V2&D 784'Z[7W<?X;O83 b b \ D]DE6F2&D 784'DE9^3Q b N O8DHQ;RS6&DHGY2&3V2&D 784ZY7P<?X;O83 b f RHD2&D _ 6&4+2 Milou strts with the initil stte: l. b b N b b b f ] i ce so he hs the intention to crry out his usul tsk of testing crisps, nd believes ll is well with the world. Hving no possible ction, Milou s met-level control function indictes he should deliberte, nd he genertes new set of intentions i c i f i to chieve the intention of testing the crisps he must first fetch them nd the first step in this fetching is to go to by the short route. t this point decides to ct, clls the ction selection function ), nd ) selects ction $. s result, Milou strts to go to. Midwy through this ction, Milou relises this ction hs filed becuse the door to is closed tht is, he revises his beliefs to get E` b b N b b =` b f, nd then decides to deliberte. This delibertion genertes new set of intentions i c i f i b. then chooses to ct, ) selects $ b, nd Milou strts to execute $ b. When this ction is complete, there is, once gin, no ction to execute, nd the met-level controller once more decides to deliberte. The reson for stepping through the exmple like this is to highlight three prticulr issues tht need to be solved in order to use DI systems, which work t precisely this kind of level of detil, with mobile robots. First, there is the issue of moving from intentions to ctions. lthough our description is little bstrct, ssuming tht there is single ction to chieve ech intention, it is close to the relity of implemented DI systems. For instnce, PRS [6] works out how to chieve intentions by pulling pre-compiled plns from pln librry. Mobile robots will require rther more sophisticted plnners, in prticulr plnners which cn pln robustly under the considerble uncertinty tht rel world mobile robots re subject to. Second, there is the whole issue of when to deliberte s ginst when to ct. Experimentl work on the problem [9, 16] hs concentrted on the reltionship between the speed of chnge of n environment nd the

3 frequency of redelibertion. Our sitution is more subtle becuse considerble effort cn be expended in trying to chieve n intention tht is no longer chieveble (like trying to pss through closed door, outside which Milou will circle forever), it is necessry to be ble to detect the filure of pln during execution. Third, there is the need to hndle uncertinty in Milou s view of the world. While the forml model ssumes boolen beliefs either Milou believes the door to is open or he believes it is closed the relity is more complex. ll Milou will hve is degree of belief, bsed on sensor input, tht the door is open or closed. s discussed elsewhere, for exmple [1, 5, 11], hndling this uncertinty requires sophisticted models. To solve these problems we turned to the use of Sffiotti s Thinking Cp [14, 15]. 4 FROM THEORY TO PRCTICE The Thinking Cp (TC) 3 is system for utonomous robot nvigtion bsed on fuzzy logic which hs been implemented nd vlidted on severl mobile pltforms [14, 15]. The min ingredients of the TC re: librry of fuzzy behviours for indoor nvigtion, like obstcle voidnce, wll following, nd door crossing; context-dependent blending mechnism tht combines the recommendtions from different behviours into trdeoff control; set of perceptul routines, including sonr-bsed feture extrction, nd detection of closed doors nd blocked corridors; n pproximte mp of the environment, together with positioning mechnism bsed on nturl lndmrks; nvigtion plnner tht genertes behviour combintion strtegy, clled -pln, tht chieves the given nvigtion gol; nd monitor tht reinvokes the plnner whenever the current -pln is no longer dequte for chieving the current gol. For the purposes of this pper, we regrd the TC s blck box tht provides robust nvigtion service, nd tht ccepts gols of the form (goto X). There re however two chrcteristics of TC tht re importnt here. First, nvigtion gols in TC re fuzzy: in (goto X), X is fuzzy loction in the robot s mp. (More precisely, gol is formlly defined in the TC frmework s fuzzy set of trjectories.) This mens tht gol in TC cn be more or less stisfied, s mesured by degree of stisfction, rel number in the intervl b c[ de. Typiclly, this degree depends on the distnce between the robot nd the desired loction, but more complex gols my hve more complex degrees of stisfction. Second, the dequcy of the current -pln which is monitored by the TC is in fct degree of dequcy, gin mesured by number in b cy dle. This degree of dequcy is the result of the composition of three terms: 1. degree of goodness, tht tkes into ccount the prior informtion vilble bout the environment; for exmple, -pln tht includes pssing through long nd nrrow corridor hs smll degree of goodness; 2. degree of competence, tht dynmiclly considers the truth of the preconditions of the -pln in the current sitution; for exmple, if door tht hs to be crossed is found closed this degree drops to 0; nd 3. degree of conflict, tht mesures the conflict between the behviours which re currently executing in prllel. f sffio/softwre/tc/ Figure 1. Integrtion between DI delibertor nd the Thinking Cp. Now the DI model nd the Thinking Cp represent two ends of the spectrum s fr s the mentl bilities of n utonomous robot re concerned. The TC cn construct plns to chieve single high level intention (like go to the lb ), but hs no grsp of the sequence of high level intentions necessry to crry out the robot s overll gols. In contrst, the DI model (t lest in so fr s we hve nlysed it with respect to intention reconsidertion) is only concerned with high level intentions nd whether or not they should be reconsidered s its beliefs bout the world themselves chnge. These my be combined s shown in Figure 1. The DI delibertor provides the delibertion function! in the forml model, generting high-level intentions of the type (goto X) nd sending them to the TC. (In future versions, intentions my include mnipultion or observtion ctivities.) The TC implements the ction selection function ), receiving these intentions nd considering them s gols. For ech gol, it genertes -pln ech corresponds to n ction in the forml model nd strts execution. The two components run s concurrent processes, with control cycles of 2s nd 100ms respectively. The TC lso monitors this execution, nd switches to new -pln if the current one turns out to be indequte. During execution, the TC recomputes the current degrees of stisfction nd dequcy every control cycle. These degrees re sent bck to the DI delibertor. From the point of view of the delibertor, the degree of stisfction mesures how much the current intention hs been chieved, nd the degree of dequcy mesures how much this intention is considered chievble. This informtion is thus prt of the input to!. In contrst to the stndrd DI model, however, this informtion is not given by binry vlues, but by continuous mesures mde possible by the use of fuzzy set theory in the TC. It is these indictors of the stte of the world vis à vis the current intention which help the delibertor to determine when it is pproprite to reconsider its intentions. Considering the forml model described bove, we should note tht, t the moment, the belief set is prtitioned between the DI interpreter nd the TC. In prticulr, beliefs tht re ffected by the dynmic nture of the world in this cse b, b N nd b b re stored in the TC nd updted s result of sensor redings. These beliefs re used to determine the degrees of stisfction nd dequcy. The more sttic knowledge is kept in the DI delibertor nd updted ccording to the degrees of stisfction nd dequcy. It is these mesures which, in prctice, cuse Milou to chnge from believing b f to ` b f when he finds tht ` b is true. These mesures, therefore, help to relte the sensor-derived beliefs b, b N nd b b which re stored in the TC to the intention determining belief b f which is stored in the DI system. The delibertor lso uses these vlues in two other importnt wys. First, to decide when it is time to deliberte. Two of the possible cuses tht led the delibertor to reconsider its intentions re: (i) n increse in the vlue of stisfction; nd (ii) drop in the vlue of dequcy. Second, it uses the vlues in the delibertion itself s

4 h i Test-Crisps Test-Crisps Lb Fetch-Crisps Tste-Crisps Fetch-Crisps Tste-Crisps Goto... Goto... Goto Lb... Goto... Goto... Goto Lb... Figure 2. Two intention trees for our exmple tsk. Lb Figure 4. Milou hs the intention (goto ) once more. Lb (2) Figure 3. (1) (3) Milou hs the intention (goto ), but this turns out to be difficult to chieve, nd dopts the new intention (goto ). mens of compring the vilble options. If this delibertion results in new intention being dopted, it is pssed to the TC. s we shll see below, considering degrees insted of binry vlues llows the delibertor to mke more informed decisions. 5 EXPERIMENTL RESULTS y wy of vlidtion of our pproch, we report n experiment where we execute the potto crisp scenrio in simulted environment. We hve used the Nomdic simultor, which includes simultion of the sonr sensors nd some moderte sensor nd positioning noise. This experiment is ment to illustrte the concepts nd mechnisms involved in our integrted pproch to robot delibertion nd nvigtion in resonbly relistic environment (lthough it cnnot tke the plce of rel experiments on live robot) 4. The successive phses of the simulted run re shown in Figures 3, 4, nd 5. Figure 6 shows the vlues of dequcy nd of stisfction of the currently executing intention t ech moment of the run. Initilly, the DI delibertor considers the new tsk nd decides strtegy, represented by the intention tree shown in Figure 2 (left). The detils of how this is done re not relevnt here (the dots indicte other intentions, like picking up the crisps, which we ignore); it suffices to note tht the intentions hve temporl order, which is tht of left to right depth-first trversl of the tree. The delibertor then psses the first intention (goto ) to TC, which genertes suitble -pln for it. In this cse there re two possible -plns, one for ech possible door leding to, nd the TC selects the one with the highest degree of (expected) goodness. Since the TC knows bout the low degree of trversbility of the lower corridor, 5 the selected -pln is the one tht goes through the min door of, the one on its left wll. Milou strts executing this -pln from the lower left corner, s indicted by (1) in Figure 3. When Milou rrives t this door (2), the sonrs detect tht the door is closed. Since one of the ssumptions in the -pln is tht the door We re currently in the process of implementing our integrted system on Nomd 200. Currently, this informtion is stored in the mp; in the future, the robot my cquire this knowledge during explortion. Figure 5. oth previous intentions re fulfilled, nd Milou dopts the intention (goto Lb). must be open, the degree of dequcy of this pln drops to 0 (Figure 6 t bout 20 s). The TC notices the problem, genertes new -pln tht goes through the second door, nd strts executing it. However, this -pln hs low degree of goodness becuse it includes pssing through the cluttered corridor. This cuses drop of the dequcy level to low vlue of 0.2. The DI delibertor notices this nd reconsiders its options. Since the current intention turns out to be difficult (but not impossible) to chieve, nd there is n lterntive wy to perform the tsk (Figure 2 right), the delibertor decides to switch to this lterntive nd to reverse the order of visiting the two production lines. Hence, it sends the new intention (goto ) to the TC (Figure 6 t 30 s). The TC genertes new -pln for this intention nd swps it in. Poor Milou then stops his journey to the lower corridor (point (3) in Figure 3), turns round, heds to room, nd eventully reches the collection point in front of conveyer belt. The chievement of the intention (goto ) is reflected in the rise of the stisfction level (Figure 6 t 75 s). This is noticed by the DI delibertor, which then sends the next intention to the TC: in our cse, this is gin the intention (goto ). Since the informtion bout closed doors inside the TC is trnsient, the TC gin genertes - pln for this intention which involves going through the min door. Milou finds his wy from room, but unfortuntely he finds tht the door is still closed (Figure 4). s before, the TC genertes n lterntive -pln going through the lower corridor nd strts to execute it. This produces drstic drop in the dequcy level, which is noticed by the DI delibertor (Figure 6 t 160 s). However, this time there is no lterntive option, so the delibertor decides to keep with the current intention, even though it is difficult to chieve. The nvigtion functionlities of the TC llow Milou to sfely, if slowly, get round the obstcles, nd rech the collection point in front of conveyer belt. The first two intentions re now fulfilled, nd the DI delibertor sends the lst one (goto Lb) to the TC. gin, the TC tries the min door first. This time we re lucky, since someone hs ctully opened this door, nd Milou eventully finds his wy to the lb, thus

5 j k dequcy Stisfction Time (goto ) (goto ) (goto ) no chnge (goto Lb) Time (goto ) (goto ) (goto ) no chnge (goto Lb) Figure 6. Mesures of dequcy (top) nd stisfction (bottom) sent by the TC to the delibertor during the run. The rrows indicte the delibertion points, nd the new intentions generted. completing the mission (Figure 5). 6 DISCUSSION This pper extends two lines of work; work on intention reconsidertion, nd work on integrting low-level nvigtion nd higherlevel resoning. It extends existing work on intention reconsidertion [9, 16, 17], by considering more complex environments, identifying whole rnge of issues which hve not been so pprent before. These include the need to build nd execute robust plns, the bility to detect the prtil filure of those plns, the bility to mesure the impossibility of chieving current intentions in the presence of uncertin informtion, nd the bility to use uncertin beliefs bout the environment in the delibertion process. Hving identified these issues, we hve proposed solution bsed on the integrtion of trditionl DI system with the Thinking Cp softwre. There re lredy number of proposls which use DI pproch to integrte low-level nvigtion nd higher-level resoning. For exmple, in [8, 10, 12] PRS-like systems re used to rbitrte low-level processes. Our proposl deprts from these pproches in the wy we prtition the responsibilities between the Thinking Cp nd the DI delibertion system. We rely on the underlying nvigtion bilities of the TC to tke cre of fuzzy behviour rbitrtion nd blending in sophisticted wy. nd we limit the role of the delibertion system to tke cre of higher level decisions bout which overll nvigtion gol should be pursued next. This prtition llows us to mke better use of the respective powers of the TC nd of the DI level. In prticulr, by pssing performnce mesures from the lower to the upper level we llow the ltter to tke more bstrct, yet still fully informed, decisions. There re two importnt wys in which our pproch cn be developed. First, the informtion pssed by the TC to the DI level could be much richer, including, for exmple, the resons why -pln hs (prtilly) filed, the conditions tht would increse its level of dequcy, or indictions bout the existence of lterntive -plns nd their degrees of dequcy. This would help the DI system in its delibertion (for instnce in determining whether to drop n intention or try to chieve it lter). Second, the choice of the strtegy used to decide when the DI should deliberte nd when it should let the TC do its job depends on the chrcteristics of the environment, nd it my itself be the result of nother, higher level delibertion. Including this ide in our frmework would led to tower of metcontrollers similr to the one suggested in [17]. Such n pproch would llow the robot to dynmiclly djust its policy for redelibertion if it finds tht the policy is incorrect with respect to its current environment. We re currently working on both these developments. cknowledgements: This work benefitted from discussions with Dimiter Drinkov nd ws prtilly supported by Queen Mry nd Westfield College, nd the Swedish KK Foundtion. REFERENCES [1] R. C. rkin, ehvior-sed Robotics, MIT Press, Cmbridge, M, [2] M. E. rtmn, Intentions, Plns, nd Prcticl Reson, Hrvrd University Press: Cmbridge, M, [3] M. E. rtmn, D. J. Isrel, nd M. E. Pollck, Plns nd resourcebounded prcticl resoning, Computtionl Intelligence, 4, , (1988). [4] P. R. Cohen nd H. J. Levesque, Intention is choice with commitment, rtificil Intelligence, 42, , (1990). [5] D. Drinkov nd. Sffiotti eds., Fuzzy logic techniques for utonomous vehicle nvigtion, Springer, erlin, Germny, [6] M. P. Georgeff nd F. F. Ingrnd, Decision-mking in n embedded resoning system, in Proceedings of the Interntionl Joint Conference on rtificil Intelligence, pp , Detroit, Michign, (1989). [7] M. P. Georgeff nd. L. Lnsky, Rective resoning nd plnning, in Proceedings of the Sixth Ntionl Conference on rtificil Intelligence, pp , Settle, Wshington, (1987). [8] F. F. Ingrnd, R. Chtil, R. lmi, nd F. Robert, PRS: high level supervision nd control lnguge for utonomous mobile robots, in Proceedings of the Interntionl Conference on Robotics nd utomtion, Minnepolis, MN, (1996). [9] D. Kinny nd M. P. Georgeff, Commitment nd effectiveness of situted gents, in Proceedings of the Interntionl Joint Conference on rtificil Intelligence, pp , Sydney, ustrli, (1991). [10] K. Konolige, K. L. Myers, E. H. Ruspini, nd. Sffiotti, The Sphir rchitecture: design for utonomy, Journl of Experimentl nd Theoreticl rtificil Intelligence, 9, , (1997). [11] D. Kortenkmp, P. onsso, nd R. Murphy eds., rtificil Intelligence nd Mobile Robots, MIT Press, Cmbridge, M, [12] J. Lee, M. J. Huber, E. H. Durfee, nd P. G. Kenny, UM-PRS: n implementtion of the procedurl resoning system for multirobot pplictions, in Proceedings of the I/NS Conf. on Robots in Field, Fctory, Service nd Spce, (1994). [13]. S. Ro nd M. Georgeff, Decision procedures for DI logics, Journl of Logic nd Computtion, 8(3), , (1998). [14]. Sffiotti, K. Konolige, nd E. H. Ruspini, multivlued-logic pproch to integrting plnning nd control, rtificil Intelligence, 76, , (1995). [15]. Sffiotti, E. H. Ruspini, nd K. Konolige, lending rectivity nd gol-directedness in fuzzy controller, in Proceedings of the 2nd IEEE Interntionl Conference on Fuzzy Systems, pp , Sn Frncisco, C, (1993). [16] M. Schut nd M. Wooldridge, Intention reconsidertion in complex environments, in Proceedings of Interntionl Conference on utonomous gents, (2000 (to pper)). [17] M. Wooldridge nd S. Prsons, Intention reconsidertion reconsidered, in Intelligent gents V, eds., J. P. Müller, M. P. Singh, nd. Ro, Springer-Verlg, erlin, Germny, (1999).