Who Needs Emotions The Brain Meets the Robot - Fellous & Arbib Part 18 doc

324 robots explicit, intended communication or by the intended actions they take in the world. Further emotional signals are communicated across a variety of channels, verbally and nonverbally. These channels vary in capacity, the specificity of the information effectively communicated, and the cognitive overhead in using them. A person can smile at a cute baby without much thought but may need more resources to verbally express happiness. Agent teams typically have two channels: communication and action. These dif- ferences suggest potential benefits for using emotions in pure agent teams. For instance, there might be an advantage to having agent teams communi- cate attitudinal or emotional information as well as an advantage to expos- ing this information to teammates automatically, through low-cost channels. Consider building agents so that they could not only communicate and act deliberately after an accurate and possibly computationally intensive assess- ment of the state, but also emit some low-cost emotional signal based on an approximate state assessment. For example, a robot could have hardwired circuitry that triggers light-emitting diodes that represent emotional cues like fear to indicate a state where the robot is in danger, worry to indicate low likelihood of success, and helplessness to indicate that it needs to help. These emotional cues can be computed and transmitted quickly and could result in the team being able to coordinate itself without having to wait for the accurate state estimation to be performed. If, for example, agents could use these emotional cues to determine action selection of the other agents in the team, it could result in greater synchronization and, consequently, bet- ter teamwork. EXPERIMENTAL ILLUSTRATION In this section, as an illustration of the effect of emotions on multiagent team- work, we demonstrate how the allocation of roles in a team is affected by emotions like fear. Our approach is to introduce an RMTDP (Nair, Tambe, & Marsella, 2003) for the team of agents, then to model the agents such that their emotional states are included. We now demonstrate how emotions can affect decision making in a team of helicopters. To this end, recall the RMTDP analysis of TOPs men- tioned above. The emotional state of the agent could skew how the agent sees the world. This could result in the agent applying different transition, observation, or reward functions. In this discussion, we will focus on how fear may affect the reward function used in the RMTDP. For instance, in a fearful state, agents may consider the risk of failure to be much higher than in a nonfearful state. In the helicopter domain, such agents might the role of emotions in multiagent teamwork 325 penalize heavily those states where a helicopter crashes. We now demon- strate how such a change in the emotional state of the agents would affect the best role allocation. We consider a team of six helicopters and vary the number of agents that fear losing a helicopter to enemy fire. These agents would place a heavy penalty on those states where one or more helicopter crashed. Figure 11.4a,b shows the number of scouts allocated to each route (X-axis) as we vary the number of fearful agents in the team (Y-axis) from none to all six for two different penalties for helicopter crashes. In Figure 11.4a, when all the agents were fearless, the number of scouts sent out was three, all on route 2; how- ever, when fearful agents were introduced, the number of scouts sent out changed to four, also on route 2, because the team was now prepared to lose out on the chance of a higher reward if they could ensure that each scout that was sent out would be safer. In Figure 11.4b, we reduced the amount of penalty the agents ascribed to a helicopter crash. When fearful agents were introduced, the number of scouts remained unchanged but the scouts now used route 1, a safer albeit longer route, instead of route 2, which was more dangerous but allowed the mission to be completed more quickly. Thus, with the introduction of fear, we found that the team’s decision-making behav- ior changed such that the members either deployed more scouts or assigned the scouts to a safer route. Figure 11.4. Role allocations in fearful teams with different reward functions. Role allocations for reward function. (a) Increasing the number of fearful agents results in more scouts being sent together to increase the safety of the scouting team. (b) Increasing the number of fearful agents results in moving scouts from a shorter but more risky route to a longer but safer route. 0. 1. 2. 5 0 5 1 5 2 5 3 3. 012345 6 Scouts on Route 1 Scouts on Route 2 Scouts on Route 3 Number of fearful agents Number of scouts on each route 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 012345 6 Number of fearful agents Number of scouts on each route Scouts on Route 1 Scouts on Route 2 Scouts on Route 3 ab 326 robots Although, the emotion “fear” was modeled simply as a penalty for states where a helicopter crashes, the purpose of the experiment was simply to show that emotional response affects what the team perceives is its best al- location. In order to evaluate teams where emotions are represented more realistically , we would need the following: • A more realistic model of how an agent’s emotional state would change based on new percepts. This model of how the emotional state transitions can be incorporated as part of the transition function in the RMTDP model in order to evaluate the team’s performance in the presence of emotion. • A more realistic model of how humans (which the agents are simulating) would respond based on their emotional state. This would form part of the TOP where the individual agent’s action selection is specified. Both the model of how emotional state changes as well as the model of human behavior in the presence of emotion should ideally be informed by human behavior in such task domains. CONCLUSION This chapter represents the first step in introducing emotions in multiagent teamwork. We examined the role of emotions in three different kinds of team: first, in teams of simulated humans, introducing emotions results in more believable agent behavior and consequently better simulations; second, in virtual organizations, where agents could simulate emotions to be more believable and engaging to the human and anticipate the human’s needs by modeling the human; and third, in pure agent teams, where the introduc- tion of emotions could help bring in the same advantages that emotions bring to human teams. Teams of simulated agents and mixed human–agent teams can greatly benefit with computational models of emotion. In particular, to evaluate and improve such teams, we would need the following: • A model of how an agent’s emotional state would change based on new percepts • A model of how humans would respond based on their emotional state Acknowledgment This research was supported by grant 0208580 from the National Science Foundation. the role of emotions in multiagent teamwork 327 References André, E., Rist, T., Mulken, S. V., & Klesen, M. (2000). The automated design of believable dialogues for animated presentation teams. In J. Cassell, J. Sullivan, S. Prevost, & E. Churchill (Eds.), Embodied conversational agents (pp. 220–255). Cambridge, MA: MIT Press. Bernstein, D. S., Zilberstein, S., & Immerman, N. (2000). The complexity of de- centralized control of MDPs. In C. Boutilier & M. Goldszmidt (Eds.), Proceed- ings of the 16th Conference in Uncertainty in Artificial Intelligence (pp. 32–37), Stanford University. Stanford, CA: Morgan Kaufmann. Boutilier, C. (1996). Planning, learning and coordination in multiagent decision pro- cesses. In Y. Shoham (Ed.), Proceedings of the Sixth Conference on Theoretical Aspects of Rationality and Knowledge (pp. 195–210). De Zeeuwse Stromen, The Netherlands: Morgan Kaufmann. Campos, J., & Sternberg, C. (1981). Perception, appraisal and emotion: The onset of social referencing. In M. Lamb & L. Sherrod (Eds.), Infant social cognition (pp. 273–314). Hillsdale, NJ: Erlbaum. Cavazza, M., Charles, F., & Mead, S. J., (2002). Interacting with virtual characters in interactive storytelling, In Proceedings of the First International Joint Confer- ence on Autonomous Agents and Multiagent Systems (pp. 15–19). Bologna, Italy: ACM. Chalupsky, H., Gil, Y., Knoblock, C., Lerman, K., Oh, J., Pynadath, D., Russ, T., & Tambe, M. (2001). Electric elves: Applying agent technology to support human organizations. In H. Hirsh & S. Chien (Eds.), Proceedings of the Thirteenth Inno- vative Applications of Artificial Intelligence Conference. Seattle, WA: AAAI. Cohen, P. R., & Levesque, H. J., (1991). Teamwork. In Noûs, 25(4), 487–512. Darwin, C. (1998). The expression of emotions in man and animals (3rd ed.). New York: Oxford University Press. (Original work published 1872) Ekman, P. (2001). Telling lies: Clues to deceit in the marketplace, politics and mar- riage. New York: Norton. Elliott, C. (1992). The Affective Reasoner: A Process Model of Emotions in a Multi- Agent System. Evanston, IL, Northwestern University Institute for the Learn- ing Sciences. Dissertation. El Nasr, M. S., Yen, J., & Ioerger, T. (2000). Flame: Fuzzy logic adaptive model of emotions. Journal of Autonomous Agents and Multiagent Systems, 3, 219–257. Frijda, N. (1987). Emotion, cognitive structure, and action tendency. Cognition and Emotion, 1, 115–143. Goleman, D. (1995). Emotional intelligence. New York: Oxford University Press. Grosz, B., & Kraus, S. (1996). Collaborative plans for complex group action. Arti- ficial Intelligence, 86, 269–357. Hatfield, E., Cacioppo, J., & Rapson, R. (1994). Emotional contagion, Cambridge: Cambridge University Press. Howard, R. A. (1960). Dynamic programming and Markov processes. Cambridge, MA: MIT Press. Hunsberger, L., & Grosz, B. (2000). A combinatorial auction for collaborative 328 robots planning. In Proceedings of Fourth International Conference on Multiagent Sys- tems (ICMAS-2000) (pp. 151–158). Boston: IEEE Computer Society. Izard, C. (1993). Four systems for emotion activation: Cognitive and noncognitive processes. Psychological Review, 100, 68–90. Jennings, J. (1990). Teamwork: United in victory. Englewood Cliffs, NJ: Silver Burdett Press. Jennings, N. (1995). Controlling cooperative problem solving in industrial multi- agent systems using joint intentions. Artificial Intelligence, 75, 195–240. Katzenbach, J., & Smith, D. K. (1994). The wisdom of teams. New York: Harper Business. Kitano, H., Asada, M., Kuniyoshi, Y., Noda, I., & Osawa, E. (1997). RoboCup: The robot world cup initiative. In Proceedings of First International Conference on Autonomous Agents (Agents ’97) (pp. 340–347), Marina del Rey, CA, Feb 5–8. New York: ACM Press. Kitano, H., Tadokoro, S., & Noda, I. (1999). RoboCup-Rescue: Search and rescue for large scale disasters as a domain for multiagent research. In Proceedings of IEEE Conference on Systems, Men, and Cybernetics (SMC-99). Tokyo, Japan: IEEE System, Man, and Cybernetics Society. Lazarus, R. (1991). Emotion and adaptation. New York: Oxford University Press. Lisetti, C. L., & Schiano, D. (2000). Facial expression recognition: Where human– computer interaction, artificial intelligence, and cognitive science intersect. Pragmatics and Cognition, 8 185–235. Marsella, S., & Gratch, J. (2002). A step toward irrationality: Using emotion to change belief. In Proceedings of First International Joint Conference on Autono- mous Agents and Multi-agent Systems (AAMAS-02) (pp. 334–341). Bologna, Italy: ACM. Marsella, S., & Gratch, J. (2003). Modeling coping behavior in virtual humans: Don’t worry, be happy. In Proceedings of Second International Joint Conference on Autonomous Agents and Multi-agent Systems (AAMAS-03) (pp. 313–320). Melbourne: ACM. Marsella, S., Johnson, W. L., & LaBore, C. (2000). Interactive pedagogical drama. In Proceedings of Fourth International Conference on Autonomous Agents (ICMAS- 2000) (301–308). Barcelona, Spain: ACM. Moffat, D., & Frijda, N. (1994). Where there’s a will there’s an agent. In Proceed- ings of Workshop on Agent Theories, Architectures and Languages (ATAL-95) (pp. 245–260). Amsterdam: Springer. Nair, R., Tambe, M., & Marsella, S. (2003). Role allocation and reallocation in multiagent teams: Towards a practical analysis. In Proceedings of Second In- ternational Joint Conference on Autonomous Agents and Multi-agent Systems (AAMAS-03) (pp. 552–559). Melbourne: ACM. Oatley, K. (1992). Best laid schemes: The psychology of emotions. Cambridge: Cam- bridge University Press. Ortony, A., Clore, G., & Collins, A. (1988). The cognitive structure of emotions. Cambridge: Cambridge University Press. Picard, R.W. (1997). Affective computing. Cambridge, MA: MIT Press. the role of emotions in multiagent teamwork 329 Pynadath, D. V., & Tambe, M. (2002). The communicative multiagent team deci- sion problem: Analyzing teamwork theories and models. Journal of Artificial Intelligence Research, 16, 389–423. Reilly, N. (1996). Believable Social and Emotional Agents. Pittsburgh: Carnegie Mellon University. Dissertation. Rickel, J., Marsella, S., Gratch, J., Hill, R., Traum, D., & Swartout, W. (2002). Toward a new generation of virtual humans for interactive experiences. IEEE Intelligent Systems, 17(4), 32–38. Scerri, P., Pynadath, D. V., & Tambe, M. (2002). Towards adjustable autonomy for the real world. Journal of Artificial Intelligence Research, 17, 171–228. Scherer, K. (1984). On the nature and function of emotion: A component process approach. In K. R. Scherer & P. Ekman (Eds.), Approaches to emotion (pp. 293– 317). Hillsdale, NJ: Erlbaum. Schut, M. C., Wooldridge, M., & Parsons, S. (2001). Reasoning about intentions in uncertain domains. In Proceedings of European Conference on Symbolic and Quantitative Approaches to Reasoning with Uncertainty, Toulouse, France. Lecture Notes in Computer Sciences, 2143, 84–85. Smith, C. A., & Scott, H. S. (1997). A componential approach to the meaning of facial expressions. In J. A. Russell & J. M. Fernández-Dols, (Eds.), The psychol- ogy of facial expression (pp. 229–254). Paris: Cambridge University Press. Tambe, M. (1997). Towards flexible teamwork. Journal of Artificial Intelligence Re- search, 7, 83–124. Tidhar, G. (1993). Team-oriented programming: Social structures. Technical report 47. Melbourne, Australia: Australian A.I. Institute. Velásquez, J. (1998). When robots weep: Emotional memories and decision-mak- ing. In Proceedings of Fifteenth National Conference on Artificial Intelligence (AAAI- 98) (pp. 70–75), Madison, WI: Cambridge, MA: MIT Press. Wooldridge, M. (2000). Intelligent agents. In G. Weiss (Ed.), Multiagent systems: A modern approach to distributed AI (pp. 27–78). Cambridge, MA: MIT Press. This page intentionally left blank PART IV CONCLUSIONS This page intentionally left blank Beware the Passionate Robot michael a. arbib 12 The warning, “Beware the Passionate Robot,” comes from the observa- tion that human emotions sometimes have unfortunate effects, raising the concern that robot emotions might not always be optimal. However, the bulk of the chapter is concerned with biology: analyzing brain mecha- nisms for vision and language to ground an evolutionary account relat- ing motivational systems to emotions and the cortical systems which elaborate them. Finally, I address the issue of whether and how to char- acterize emotions in such a way that one might say that a robot has emotions even if they are not empathically linked to human emotions. A CAUTIONARY TALE On Tuesday, I had an appointment with N at 3 P.M., but when I phoned his secretary at 2:45 to check the place of the meeting, I learned that she had forgotten to confirm the meeting with N. I was not particularly upset, we rescheduled the meeting for 4 P.M. the next day, and I proceeded to make contented use of the unexpected free time to catch up on my correspondence. On Wednesday, I decided in midafternoon to put together a chart to discuss with N at our meeting; but the printer gave me some problems, and it was already after 4 when I left my office for the meeting, feeling somewhat flustered but glad that I had a useful [...]... does he not want these crates any more? But none of these is the right thought The right thought to think is: How does one use the crates to reach the bananas? Sultan drags the crates under the bananas, piles them one on top of the other, climbs the tower he has built, and pulls down the bananas He thinks: Now will he stop punishing me? At every turn Sultan is driven to think the less interesting... another example of the competition and cooperation that is so distinctively the computing style of the brain (Arbib, 1989) Note the role here of social norms in judging the behavior to be inappropriate with the concomitant emotion of shame, providing the motivation to take a beware the passionate robot 341 course of action (apology) that will make amends for the inappropriate behavior; note that the. .. useful information from a particular type of sensory input These basic systems make data available which can provide the substrate for the evolution of higher-level systems to extract new properties of the sensory input The higher-level systems then enrich the information environment of the basic systems by return pathways The basic systems can then be adjusted to exploit the new sources of information... at the motivational systems which ground the emotions HUGHLINGS JACKSON: AN EVOLUTIONARY FRAMEWORK I now offer a general framework for the study of the evolution of brain mechanisms which will inform the following two sections Hughlings Jackson was a 19th century British neurologist who viewed the brain in terms of levels of increasing evolutionary complexity (Jackson, 187 8–79) Influenced by the then... screaming out the student’s shortcomings in emotional tones laced with in- beware the passionate robot 335 vective and carrying out the electronic equivalent of tearing up the student’s papers in a rage, then where would the benefit lie? One might argue that even though such outbursts are harmful to many children, they may be the only way to “get through” to others; but if this is so, and the production... strategy, it may be debated whether the computer tutor really has emotions or is simply “simulating the appearance of emotional behavior”—a key distinction for the discussion of robot emotions We will return to these questions later (see Emotion without Biology, below) To complement the above account of my own tangled emotions on one occasion, I turn to a fictional account of the mental life of a chimpanzee... primatologist who supports such “anthropomorphism”), but my point here is to stress a “two-way reductionism” (Arbib, 1985; Arbib & Hesse, 1986) which understands the need to establish a dialog between the formal concepts of scientific reductionism and the richness of personal experience that drives our interest in cognition and 336 conclusions emotion How can we integrate the imagination of the novelist with the. .. to what extent the “overthe-top” level of annoyance here labeled “fury” was targeted rather than diffuse It was not targeted at my secretary the recollection of her absence served to explain why I had not received the message (it had been left on the voicemail, which she would normally relay to me), not to blame her for being away It was the cancellation of the meeting, not the loss of the message, that... and with my thoughts dominated by this stew of emotions, I decided to return to N’s office Once there, I apologized to the secretary for my discourtesy, explained the annoyance of the double cancellation, set a new appointment, made a feeble joke in the form of a threat about the dire consequences of another cancellation, and then returned to my office The physiological effects I had felt just a few... perspective offered later, under From Drives to Feelings These insights will then serve to anchor a fresh look at the issue of robot emotions under Emotion without Biology Following each excerpt from the narrative, I offer a sequence of mental states and events, without teasing out the overlapping of various segments The notion x will denote the experience of the emotional state x, by which I mean a reportable . PART IV CONCLUSIONS This page intentionally left blank Beware the Passionate Robot michael a. arbib 12 The warning, “Beware the Passionate Robot, ” comes from the observa- tion that human emotions. debated whether the computer tutor really has emotions or is simply “simulating the appearance of emotional behavior”—a key distinc- tion for the discussion of robot emotions. We will return to these. provides another example of the com- petition and cooperation that is so distinctively the computing style of the brain (Arbib, 1989). Note the role here of social norms in judging the behavior