Modeling User Perception of Interaction Opportunities for Effective Teamwork

Transcription

1 Modeling User Perception of Interaction Opportunities for Effective Teamwork Ece Kamar, Ya akov Gal and Barbara J. Grosz Scool of Engineering and Applied Sciences Harvard University, Cambridge, MA USA {kamar, gal, Abstract Tis paper presents a model of collaborative decision-making for groups tat involve people and computer agents. Te model distinguises between actions relating to participants commitment to te group and actions relating to teir individual tasks, uses tis distinction to decompose group decision making into smaller problems tat can be solved efficiently. It allows computer agents to reason about te benefits of teir actions on a collaboration and te ways in wic uman participants perceive tese benefits. Te model was tested in a setting in wic computer agents need to decide weter to interrupt people to obtain potentially valuable information. Results sow tat te magnitude of te benefit of interruption to te collaboration is a major factor influencing te likeliood tat people will accept interruption requests. Tey furter establis tat people s perceived type of teir partners (weter umans or computers) significantly affected teir perceptions of te usefulness of interruptions wen te benefit of te interruption is not clear-cut. Tese results imply tat system designers need to consider not only te possible benefits of interruptions to collaborative uman-computer teams but also te way tat suc benefits are perceived by people. I. INTRODUCTION Tis paper considers collaborative decision-making in eterogeneous groups of computer agents and people. Participants of a collaborative activity work togeter towards satisfying a joint commitment, but tey also adopt and care about teir individual goals. To be successful in suc settings, agents need to reason about te effects of teir actions on te collaboration and te way oter participants perceive tese actions. Tis paper proposes a new model for collaborative decision-making tat significantly reduces te complexity of making joint decisions. Te model distinguises between tose situations in wic agents actions affect te collaboration, and tose in wic teir actions affect only teir individual tasks. It decomposes te collaboration into smaller interacting subproblems tat can be analyzed independently and combined to capture te collaborative utility of an action. Wen people participate in collaborative activities wit computer agents it is necessary for te agents to reason about te ways in wic people perceive te utility of te collaboration and its constituent actions. We empirically investigate te mismatc between te actual utility of an action in a collaborative context and people s perception of it, exploring te different factors tat may influence people s perception of tis utility. Te failure to consider tis mismatc may cause a person to reject a valuable interaction opportunity, tereby turning wat could ave been a beneficial interaction for te collaboration into a performance degrading disturbance. Our investigations use interruption management as an example of a decision making capability needed for collaborative activities in wic agents are distributed, conditions may be rapidly canging and decisions are made under uncertainty. Interruptions are important for effective collaborative work, because agents often possess information required by oters on teir team. However, interruptions are inerently disruptive. If tey are not managed and timed properly, tey may negatively affect te emotional state and awareness of te user and may reduce te overall task performance of te user and te system [1]. For example, a writer s collaborative assistant tat autonomously searces for bibliograpical and citation information needs to know wen to ask user for information and ow to time requests [2]. If te assistant continuously asks weter to cite eac paper tat meets a user s keywords and commands it will disrupt te user s writing process. Our decoupling model for collaborative decision making, called DECOP, syntesizes tecniques from decision teory and computer science, but adapts tem to collaborative contexts. DECOP takes into account te costs and benefits to all participants, weter uman or computer agent, so tat decisions to interrupt are based on te collaborative benefit to te group. Unlike previous models of interruption management, DECOP also reasons about te possible mismatc between a computer s estimate of te utility of an interruption and a person s perception of it. It focuses on determining te factors tat influence people s perception of interruptions, and teir tendency to accept or reject tem wen tey are generated by a computer system. We constructed a computer agent based on DECOP model in an empirical collaborative setting to investigate te way people perceive interruptions. We analyzed te effect of tree factors on uman perception of te usefulness of interruption requests: te magnitude of te interruption utility, te timing of interruptions and te perceived type of te partner (uman or computer agent). Te results revealed tat te magnitude of te utility of interruption is te major factor affecting te likeliood tat people will accept interruptions. Te results also indicate tat te perceived partner type and cost of te interruption to te subject affect people s perception of te usefulness interruptions wen te utility gain is less clear-cut.

2 II. RELATED WORK Computing joint strategies for multi-agent decision-making problems under uncertainty as sown to be infeasible for realistic problems [3]. In tis work, we introduce a new approac for efficient collaborative decision making in te spirit of near-decomposable models [4]. Te remainder of tis section compares our approac to te previous work in te domain of interruption management. A key aspect of reasoning about interruptions in collaborative settings is te ability to accurately estimate te costs and benefits of te interruption to all parties so tat te outcome of te interruption positively affects group task outcomes. Previous work on adjustable autonomy identifies te points at wic it is most suitable to initiate interactions wit a person, but does so witout relating tis decision to a user s mental state or te task being performed. Interruptions are driven solely by system needs and managed based on benefit to te system [5]. Prior work on interruption management as addressed user needs, but as focused mostly on te effect an interruption as on a person s cognitive state, rater tan te benefit of interruptions to collaborative activities [6]. Few models ave combined tese two aspects into an integrated decision making mecanism [7], and none ave done so in te kinds of rapidly canging domains of uncertainty we consider. Wile tere as been significant work on mixed-initiative system design, tere as been little empirical work on ow people perceive interruption utilities and make interruption decisions in uman-computer interaction settings. Avraami et al. [8] investigated te differences between a person s self report of interruptibility and oter people s predictions about tat person s interruptibility. However, tis work considered face to face uman interaction, rater tan uman-computer interaction. Gluck et al. [9] focused on designing notification metods to increase uman perception of utility wereas Bunt et al. [10] sowed tat displaying system rationale to people may induce a person to trust a computer system more. III. THE INTERRUPTION GAME Tis section describes a game designed for investigating te interruption management problem in a setting tat does not require sopisticated domain expertise. Te interruption game involves two players, referred to as te principal and te agent. Eac player needs to complete an individual task but te two players scores depend on eac oter s performance making tis a collaborative endeavor. Te game is played on a board of 6x6 squares. Eac player is allocated a starting position and a goal position on te board. Te game comprises a fixed, known number of rounds. At eac round, players advance on te board by moving to an adjacent square. Te players goals move stocastically on te board according to a Gaussian probability distribution centered at te current position of te player. 1 Players earn 10 points in te game eac time tey move to te square on wic 1 Te movement of te goal is restricted in tat it does not move closer to te position of te player. Fig. 1. Game Screen-sot: me is te principal player, smiley is te agent player, G me is te principal s goal, G smiley is te agent s goal. Te degree to wic eac square is saded represents te agent s uncertainty about its goal. Dark squares imply iger certainty. teir assigned goal is located, and te goals are reassigned to random positions on te board. Players can see teir positions and te goal location of te oter player, but tey differ in teir ability to see teir own goal location: Te principal can see te location of its goal trougout te game, wile te agent can see te location of its goal at te onset of te game, but not during consecutive rounds. At any round, te agent can coose to interrupt te principal and request te current location of its goal. Te principal is free to accept or reject an interruption request. If te principal rejects te interruption request, te players continue moving. If te interruption is accepted by te principal player, te location of te agent s goal in te current round (but not in consecutive rounds) is automatically revealed to te agent. Tere is a joint cost for revealing tis information to te agent, in tat bot participants will not be able to move for one round. Te game scenarios used in te empirical evaluation are simplified to allow a single interruption troug te game. Te rules of te game also require te agent to provide te principal wit its belief about te location of its goal. Tis information may influence te principal s decision about weter to accept an interruption. A snapsot of te game from te perspective of te principal player is sown in Figure 1. Te rules of te game provide incentives to players for reacing teir goals as quickly as possible and interruptions initiated by te agent are critical determinants of players performance. Te agent s uncertainty about te location of its own goal increases over time, and its performance depends on successfully querying te principal and obtaining te correct position of its goal. Te game is collaborative in tat te score for eac player depends on a combination of its own performance and te performance of te oter player. Te players sare a joint

3 score function tat is te cumulative score of bot players. An interruption is potentially beneficial for te individual performance of te agent, wo can use tis information to direct its movement, but it only induces te principal s performance negatively. Providing tis information is costly for bot players. Wen te agent deliberates about weter to ask for information, or wen te principal deliberates about weter to reveal te information to te agent, te players need to weig te trade-offs associated wit te potential benefit to te agent player wit te detriments to teir individual performance in te game. Te success of bot players in te game depends on te agent s ability to estimate te collaborative value of interruption at eac point in te game and use tat information to coose wen to interrupt te principal. Te interruption game is not meant to be a complete model of any specific domain or application. Its purpose is to provide a simple setting in wic to study te factors tat influence interruption management in collaborative settings. It provides a setting tat is analogous to te types of interactions tat occur in collaborative settings involving a mixed network of computer agents and people. For example, te principal player in te interruption game may represent te user of a collaborative system for writing an academic paper, and te agent may represent te collaborative assistant responsible for obtaining bibliograpical data. Wile bot of te participants sare a common goal of completing a document, eac of tem must work independently to complete its individual task, suc as composing paragraps or searcing for bibliograpical information. Tis aspect is represented in te interruption game by assigning an individual goal for eac player. Te movement of tese goals on te board corresponds to te dynamic nature of tese tasks. For example, te user may not know wat to write next, and te system may ave uncertainty about searc results. Te agent s lack of information about its own goal location in te game corresponds to te uncertainty of a system about te preferences and intentions of its user, suc as wic bibliograpical information to include in te paper. Te ability to query te user for keywords and to coose among different bibliograpies provides te system wit valuable guidance and direction. It may, owever, impede te performance of bot participants on teir individual tasks, because te system needs to suspend its searc for bibliograpical data wen it queries te user, and te user may be distracted by te query. Tis dynamic cost of interruption represents te costs incurred to bot users and computer agents due to task switcing and task recovery for initiating and responding to an interruption. IV. MODELING INTERRUPTION OPPORTUNITIES In tis section we formalize te interruption game as a multi-agent decision-making problem under uncertainty, and we provide efficient metods for computing players estimates of te benefit of interruptions in te game. Tese models are not meant to predict ow people play te interruption game or respond to interruption requests in general. Rater, tey provide a way to compute a teoretical baseline, wic is a fully rational computational estimate for te value of interruption in te game. In te empirical section, we use tis baseline to enable empirical analysis of uman beavior in mixed-initiative settings. Tis analysis allows te study of te efficacy of tese models wen tey are used by computers to interact wit people in te game under various experimental conditions. Te interruption game can be modeled as a Decentralized Markov Decision Process (Dec-MDP) [11], a formalism for multi-agent planning tat captures te collaborative nature of te interruption game and its associated uncertainty. A Dec-MDP includes a set of states wit associated transition probabilities, a set of actions and observations for eac agent, and a joint reward function. A solution of a Dec-MDP is an optimal joint policy for all agents tat is represented as a mapping from states to actions. To model te interruption game, te state space of te Dec- MDP will combine all of te information relating to te tasks of bot players, including teir positions on te board, te positions of teir goals, te current round and te belief of te agent about its own goal position. Te solution of te Dec- MDP assigns a policy to te agent tat initiates interruption requests wen tey are expected to result in a benefit to bot players according to te joint reward function and assigns a policy for te principal to accept interruption requests tat ave actual positive benefit. Unfortunately, finding optimal solutions to Dec-MDPs is NEXP-complete [3]. Te size of te state space makes it infeasible to compute te complete joint policy for bot players in te interruption game. However, our goal is not to exaustively compute optimal policies in te interruption game, but to be able to generate interruptions wen tey are perceived to be beneficial to te collaboration. We ypotesized tat suc interruptions would be likely to be accepted by people. Our novel model for decision making, called DECOP, exploits an important caracteristic of tis game and many domains in wic computer agents and people work togeter: Wen players are not making or replying to interruption requests, tey are performing teir individual tasks, and eac player needs to consider only its individual score in te game. In tis case, te two tasks are essentially independent, and tey can be solved separately. As te agent can only interrupt te principal once, te expected utility of an interruption can be computed efficiently, because an interruption request will render te two tasks independent from te interruption moment until te end of te game. At eac turn, te policy for te agent is to interrupt and request information from te principal wen it is deemed beneficial for bot participants. Te next section details DECOP model tat captures te benefit of an interruption by solving te individual tasks for bot participants in te game, and combining tese solutions in order to devise strategies for interruption management in te game. A. Computing a Policy for te Principal Te principal as complete information about te game, so its task can be modeled as a Markov Decision Process (MDP). Let B denote te set of board positions; B denotes te size

4 of te game board; p B, g B are te positions of te principal and its goal respectively; m A is a movement action of te player; P (g, p, g) is te probability of te goal position moving from position g to position g wen te player is in position p. Te state space of te MDP includes every possible position of te principal and its goal at eac round. We denote SP = p, g to be te state at round and induce a state transition function T tat assigns a probability to reacing state S +1 P from SP given action m. T can be directly derived from P. Te reward function R assigns te score in te game for reacing te goal if an action transitions a player to its goal square, and 0 oterwise. Let Π P denote te optimal policy for te principal player in te game. Te value V Π P (S ) of tis policy at state SP maximizes te reward at state SP for an action m and future states given te transition probability function, V Π P (S P ) = max m [R(S P, m) + S +1 T (S +1, m, S P ) V Π P (S +1 P )] (1) We compute te optimal policy and its value using ExpectiMax searc. In tis process we grow a tree wit two types of nodes, decision nodes and cance nodes. Tere is a decision node for eac state in te MDP, and eac cild of a decision node is labeled wit a movement action for te principal. Cance nodes represent moves of nature, and eac cild of a cance node is labeled wit a possible movement of te goal of te principal, and is assigned a probability according to te transition function. Wen traversing te tree, we recursively compute a value for eac cance node tat is a weigted average of te value of eac of its cildren according to its probability. We compute a value for eac decision node by coosing te cild wit te maximal value, and select tat action. Wit memoization, te number of nodes generated by te searc is bounded by B 2 H, wic is polynomial in te number of rounds in te game. B. Computing a Policy for te Agent Te agent cannot observe te position of its goal on te board, and witout interrupting te principal it receives no information relating to tis position. We model its task as a No Observation Markov Decision Process (NOMDP), wic is a special case of an MDP wit no observations. Te state space for tis model includes te position l of te agent on te board, its belief b B over its goal position, and current turn. We denote te state for te agent as SA = l, b. As we are modeling te agent s individual task, rater tan its interaction wit te principal, we leave out te interruption action and use te set of actions A and reward function R identical to te ones described for te principal player. Te agent updates its belief b to b after eac turn according to te goal movement distribution P as follows: c B, b (c ) = c B b(c) P (c, l, c) (2) Te value of an optimal policy for te agent Π A at state SA = l, b can be computed using Equation 1, substituting Π A for Π P and S A for S P. Because te belief of te agent about its goal position is incorporated into te state space, tere are an infinite number of states to consider, and using ExpecitMax in a straigtforward fasion is not possible. However, applying te belief update function after eac turn, only a small number of states turn out to be reacable. Te deterministic belief update function maps eac combination of states wit full information (i.e., states in wic te agent knows te correct position of its goal) and te number of turns since full information to a single belief state, tus to a single state. As a result, we can grow te searc tree on te fly, and only expand tose states tat are reacable after eac turn. Memoization is not possible in tis tecnique, and tus te complexity of te complete searc is exponential in te lengt of te orizon. C. Computing te Benefit of Interruption To compute te benefit of an interruption, its effect on bot te agent s and te principal s individual performance must be taken into account. It is te aggregate of tese two effects tat determines te utility of interruption. An interruption is initiated by te agent, but it is only establised if te principal accepts it. Te effect of an interruption on te individual game play of a player is te difference of te values of two states; one in wic an interruption is establised, and oter in wic it is not. Given te principal and its goal are located on squares p and g respectively in game round, let EUP NI(S P = p, g ) denote te expected utility of te principal wen it is not interrupted, and pursues its individual task. Tis is equal to te value for te principal of carrying out its individual task as sown in Equation 1. Tus we ave, EU NI P (S P ) = V Π P (S P ) (3) For te agent tat does not get to observe its own goal position, let EUA NI(S A = l, b ) denote te expected utility of te agent wen it is not interrupted, and pursues its individual task. Tis is te value to te agent of carrying out its individual task: EU NI A (S A) = V Π A (S A ) (4) Let EUP I (S P = p, g ) denote te expected utility for te principal wen it accepts an interruption. If te principal player is interrupted, it cannot move for one round, but its goal may move stocastically according to te probability distribution P. We denote te new goal position as g +1. Given our constraint tat tere can only be one interruption made in te game, te benefit of interruption for te principal is te expected value of its individual task in future rounds, for any possible position of its goal. Formally, te utility of interruption for state SP, denoted EU P I (S P ) is computed as follows: EU I P (S P ) = g +1 B P (g +1, p, g) V Π P (S +1 P = p, g +1 ) (5)

5 If te agent successfully interrupts te principal, te principal will reveal te position of te agent s goal. Te agent will update its belief about its goal position in te following round, and use tis belief to perform its individual task in future rounds. However, wen it deliberates about weter to interrupt in te current round, it needs to to sum over every possible position of its unobserved goal, according to its belief about te goal location. Let SA = l, b be te current state of te agent, including its position on te board and belief about its goal position. Let g denote te current position of its goal. Te expected value for interruption for te agent is denoted EUA I and is computed as follows: EU I A(S A) = g B b(g) V Π A (S +1 A = l, b ) (6) Here, b refers to te belief state of te agent in wic probability 1 is given to g, te true position of its goal as revealed by te principal, and updated once to reflect te stocastic movement of te goal in turn. D. Deciding Weter to Interrupt By combining te expected values of te principal and agent players wit and witout interruption, it is now possible to compute te agent s estimate of te benefit of an interruption. We denote EBI P (SP ) to be te expected benefit of interruption for te principal given SP, wic is simply te difference in utility of te principal between accepting and interruption and carrying out its individual task. EBI P (S P ) = EU I P (S P ) EU NI P (S P ) (7) Te expected benefit of interruption for te agent is denoted EBI A (SA ) and is computed similarly: EBI A (S A) = EU I A(S A) EU NI A (S A) (8) Te interruption game is collaborative in tat te combined performance of bot participants determines teir individual scores. Te agent can observe te state SP of te principal, and for any combined state S = (SP, S A ), te agent will consider te joint expected benefit of interruption to bot participants, EBI, and coose to interrupt if tis joint benefit is positive. EBI(S ) = EBI P (S P ) + EBI A (S A) (9) Te agent cannot observe te correct position of its goal and estimates te benefit of interrupting under tis uncertainty. Tus, not every interruption initiated by te agent is truly beneficial for te team. In contrast, te principal can observe te position of agent s goal and can capture te actual benefit of te interruption, denoted ABI wit certainty. Any interruption wit positive ABI offers a positive expected benefit to te team. Te value of ABI is te sum of te individual benefits of interruption to bot te principal and te agent. Let g a be te agent s goal position, te actual benefit of interruption for bot participants given states S P and S A is ABI(S ) = EBI P (S P ) + EBI P,A (S A) (10) Here, te term EBI P,A (SA ) refers to te principal s perception of te agent s benefit from revealing te goal position g a, were l is te current position of te agent, b refers to te belief state of te agent in wic probability 1 is given to g a and updated once. EBI P,A (S A) = V Π A (S +1 A = l, b ) EU NI A (S A) (11) Te advantage of DECOP model introduced above is tat it reduces te complexity of te multi-agent decision making process to tat of two separate single agent decision making processes. Because te agent is allowed to interrupt only once during a game-play, te decoupling metod is able to accurately capture te benefit of an interruption initiated by te agent. DECOP model assumes tat principal players are fully rational and computationally unbounded. In te case of suc players, we would expect any interruption wit positive ABI to be accepted, and any interruption wit negative ABI to be rejected. However, people may not be fully rational or computationally unbounded, and we expect people s perception of te benefit of interruptions to differ from baseline values calculated by DECOP model. In te empirical investigations described in te next section, tese baseline values are compared wit te subject responses to detect te possible mismatc between a computer s estimate of te benefit of an interruption and a person s perception of it, and to identify a subset of factors tat affect te way tat umans perceive te effectiveness of interruptions. V. EMPIRICAL METHODOLOGY Tis section uses strategies derived from DECOP model for playing te interruption game to explore te way people make decisions in an empirical setting. A total of 26 subjects participated in te study. Te subjects were between ages of 19 and 46 and were given a 20 minute tutorial of te game. Subjects played between 25 to 35 games eac, and were compensated in a manner tat was proportional to teir total performance. During te empirical evaluation, all subjects were allocated to play te roles of principals, wile te role of te agent was assigned to a computer tat used te metodology described in te previous section to play te interruption game. Eac game proceeded in te manner described in Section III. In particular, te agent could not observe its own goal location, but is allowed to initiate an interruption once to acquire te correct location of its goal from te principal. At eac turn of te game, te policy of te agent is to interrupt only if te expected benefit of an interruption (EBI) is positive Interruptions were generated by te computer agents at different points in te game wit varying actual benefits, game levels and perceived partner types. We measured people s responses to tese requests given te game conditions at te time of interruptions, wic included te number of turns left to play, te positions of bot players on te board, and te agent s belief about te location of its goal. A principal player tat uses DECOP model to determine weter to accept an interruption request is perfectly rational

6 in tat is uses te collaborative benefit of interruption (ABI), given in Equation 10, as te sole factor for tis decision. However, we expected people to differ from tis rational model. Te purpose of te empirical work is to measure te extent to wic different factors in te game, suc as te collaborative benefit of interruption (ABI), te timing of interruptions and te perceived partner type, influence people s perception of interruptions. To investigate te way subject responses cange wit respect to benefit of interruptions, te game scenarios were varied to ave different ABI values. To investigate te effect of te timing of an interruption on te subjects likeliood of acceptance, we varied te level of te game tat an interruption is initiated. Lastly, we expected tat te type of agent player (weter a computer or uman) would affect te way people respond to interruption requests. For tis reason, subjects were told tey would be interacting wit a uman for some games, owever tey were always paired wit an agent 2. Subjects were given randomly generated game scenarios tat vary te actual benefit of interruption to bot participants (ABI) to cover four types of values: -1.5 (small loss), 1.0 (small gain), 3.5 (medium gain), 6.0 (large gain). Tese values represent te smallest and largest benefit values tat can be generated from interruptions wit positive expected benefit (EBI), wic is a necessary condition to initiate interruption requests by te agent player. Te levels in te game in wic interruptions occurred varied to cover te beginning of a game (level 3), te middle of a game (level 5) and te end of a game (level 7). Tere were 540 game instances played wen te perceived agent was a computer (PP:Computer) and 228 data points wen te perceived agent was a person (PP:Person). VI. RESULTS AND DISCUSSION Te following results analyze a total of 768 game instances collected in our study. Table 1 sows interruption acceptance rates for different levels and ABI values for te same game instances wen perceived partner type (PP) is person or agent. Te optimal policy for te principal player is to accept an interruption if its associated benefit (ABI) is positive and to reject oterwise. PP:Computer Level 3 Level 5 Level 7 ABI ABI ABI ABI PP:Person Level 3 Level 5 Level 7 ABI ABI ABI ABI TABLE I ACCEPTANCE RATE OF INTERRUPTIONS 2 Approval was obtained for te use of uman subjects in researc for tis misinformation. Fig. 2. Effect of interruption benefit and perceived partner type on interruption acceptance rate As te results of Table 1 sow, te utility of an interruption is te major factor influencing te probability tat an interruption will be accepted by a person. Te interruption acceptance rate increases significantly as te benefit of interruption rises from -1.5 to 1.0 (p < e 20, α=0.001) and from 1.0 to 3.5 (p=0.0013, α=0.01). However te rise from 3.5 to 6.0 does not furter improve te acceptance rate. Tese results confirm tat people were successful at perceiving interruption benefits above a certain tresold. Similarly, wen an interruption is costly for te collaboration, people are significantly more likely to reject te interruption. However, subjects varied in teir responses to interruptions offering sligtly positive gains, indicating te difficulty to estimate te benefit of interruption wen its usefulness is ambiguous. Figure 2 summarizes te acceptance rates of interruption as a function of te actual benefit of interruption and perceived partner type (person vs. computer). We divide te figure into tree regions of interruption benefits: small losses (Region 1), small gains (Region 2), and large gains (Region 3). Te analysis sows tat for large losses (Region 1) and for small gains (Region 3), canging te perceived partner type does not affect te likeliood tat te interruption will be accepted. In contrast, for interruptions offering small gains (Region 2), te acceptance rate is significantly larger if te perceived partner type is a person (p = , α = 0.001). Tis result implies tat wen te benefit of interruption is straigtforward, people do not care weter te initiator of te interruption is a person or a computer. However, for tose cases in wic te benefit of interruption is ambiguous, people are more likely to accept interruptions tat originate from oter people. Tis result suggests tat agent designers need to consider te way tey present interruptions to teir users in cases were te perceived benefit is ambiguous. It also aligns wit recent studies sowing tat mutual cooperation is more difficult to acieve in uman-computer settings as compared to settings involving people exclusively [12]. Figure 3 sows te effect of interruption timing (te level of te game) on people s acceptance rates for interruptions of small losses and small gains (Te interruption timing does not affect te acceptance rate for interruptions of large gains). We

7 strategies directly applicable for interruption management in real world domains, but rater to sow tat effective interruption management needs to consider te collaborative benefit of interruption to bot user and system, and to point system designers to te types of factors tat people consider wen tey reason about interruptions. In future work, we plan to extend te study to better understand te effects of computational and cognitive complexity on people s interruption strategies, focusing on te possible role of trust and overestimation of costs. Fig. 3. Correlation of interruption acceptance rate wit te cost of interruption to subjects (-ABI P ) for small gains and losses. expected tat interruptions occurring late in te game (i.e., wit fewer number of turns left in te game) will be more likely to be accepted wen tey incur positive benefit, and rejected wen incurring a loss. However, as sown by te Figure 3, as te game level increases, so does te acceptance rate for interruptions of bot small losses (ABI -1) and small gains (ABI 1). Tere is a significant increase in te acceptance rate wen game level increases from 3 to 5 (p=0.002, α=0.01) and from 5 to 7 (p< 10 6, α=0.001). One factor tat may explain te correlation between te acceptance rate and te game level for interruptions of small gains and losses, is te cost of interruption to te subject. As sown in Figure 3, te cost of interruption to te subject (ABI P ) decreases as game level increases. Tus, for interruptions of small gains and losses, we found tat te acceptance rate is negatively correlated wit te cost of interruption to te principal. In addition, it was revealed tat te benefit of te interruption to te principal (ABI P ) is a better predictor of te acceptance rate tan ABI A, te benefit of interruption to te agent (logistic regression SE = 0.05, R 2 = 0.19, p < 0.001). Tus, uman subjects tend to overestimate teir own benefit from interruptions as compared to te benefit for te agent. Consequently, te benefit of interruption to person may be weigted more in person decision making model tan te benefit of te interruption, and people may be more likely to accept an interruption wit low ABI P among interruptions wit identical benefit. Furter study is required to determine weter tese conjectures old and better understand te correlation of acceptance rate wit te cost of interruption to te person. Tis conjecture is supported by some subject responses to survey questions regarding teir strategies for accepting interruptions. Answers include: If te agent was in te totally wrong direction and I ad several moves left, I would allow te interruption. I always wanted te sure ting for myself. If te collaborator was way off in knowing and ad enoug moves to likely catc it after I told te location, I accepted. If it compromised my ability to get my goal, I declined. Lastly, we empasize tat tese results are a first step in understanding te uman perception of interruptions in collaborative settings. Our goal was not to design computational VII. CONCLUSION We ave presented a novel model for collaborative decisionmaking tat computes te utility of actions from te possibly different perspective of te participants of te collaborative activity. We evaluated te model in a specially designed umancomputer collaborative setting in wic computers need to manage teir interruption requests to umans. We sowed tat te actual benefit of interruptions to bot computer agents and people is te major factor affecting te likeliood tat people will accept interruption requests. However, for tose cases in wic te benefit of interruption is ambiguous, people prefer to accept tose interruptions tat originate from oter people. Our empirical studies provide a number of insigts about uman decision making in te context of uman-computer collaboration in dynamic, uncertain domains. Understanding te way tat people make interruption decisions will enable te development of better mecanisms for initiating interruptions, focusing on te interruptions tat are more likely to be accepted. Tese investigations will provide te foundation for building agents tat collaborate wit people efficiently witout overburdening tem. REFERENCES [1] P. Adamczyk and B. Bailey, If not now, wen?: te effects of interruption at different moments witin task execution, CHI, [2] T. Babaian, B. Grosz, and S. Sieber, A writer s collaborative assistant, in Proceedings of te 7t international conference on Intelligent user interfaces. ACM New York, NY, USA, 2002, pp [3] D. S. Bernstein, S. Zilberstein, and N. Immerman, Te complexity of decentralized control of markov decision processes, in UAI, [4] H. Simon, Te arcitecture of complexity, Proceedings of te American Pilosopical Society, pp , [5] P. Scerri, D. Pynadat, W. Jonson, P. Rosenbloom, M. Si, N. Scurr, and M. Tambe, A prototype infrastructure for distributed robot-agentperson teams. in AAMAS, [6] E. Horvitz and J. Apacible, Learning and reasoning about interruption, ICMI, [7] M. Fleming and R. Coen, A user modeling approac to determining system initiative in mixed-initiative ai systems, in UM, 2001, pp [8] D. Avraami, J. Fogarty, and S. E. Hudson, Biases in uman estimation of interruptibility: effects and implications for practice, in CHI, [9] J. Gluck, A. Bunt, and J. McGrenere, Matcing attentional draw wit utility in interruption, in CHI, [10] A. Bunt, J. McGrenere, and C. Conati, Understanding te utility of rationale in a mixed-initiative system for GUI customization, in UM, [11] E. Kamar and B. J. Grosz, Applying mdp approaces for estimating outcome of interaction in collaborative uman-computer settings, MSDM, [12] J. Rilling, A. Sanfey, J. Aronson, L. Nystrom, and J. Coen, neural correlates of teory of mind witin interpersonal interactions, pp