《PMFserv based crowd simulation.doc》由会员分享，可在线阅读，更多相关《PMFserv based crowd simulation.doc（14页珍藏版）》请在taowenge.com淘文阁网|工程机械CAD图纸|机械工程制图|CAD装配图下载|SolidWorks_CaTia_CAD_UG_PROE_设计图分享下载上搜索。
1、Progress to Date on the Human Performance Moderator Function Server (PMFserv) for Rapidly Generating Reusable Agents, Forces, and CrowdsJanuary 2003Barry G. Silverman, PhD.Jason B. CornwellKevin OBrienAckoff Center for the Advancement of Systems Approaches (ACASA)University of Pennsylvania, Towne 22
2、9cPhiladelphia, PA 19104-6315barrygseas.upenn.eduKEYWORDS: Software agents, culture, stress, emotion, crowd models, asymmetric adversaries, emergenceABSTRACTThis paper summarizes our latest progress on integrating human performance moderator functions (PMFs) from a range of ability, stress, emotion,
3、 decision theoretic, cultural and motivation literatures into a composable framework or server for rapidly generating synthetic asymmetric agents and scenarios. Our goal is to create the PMFserv as an open agent architecture that allows one to research and explore alternative PMFs to add realism to
4、software agents - e.g., physiology and stress, personal values and emotive states, and cultural influences. In PMFserv each agents decisions are guided by its perception of the environment and its current set of needs as determined by various reservoirs reflecting the current levels of the physiolog
5、ic, stress, emotive, and cultural PMFs. We present the progress to date on the composability of the modules, and on an environment for quickly marking up agents and world objects in terms of their need reservoir value affordances. We illustrate the approach via an example training game for counter-t
6、errorism and crowd management. Future research needs are elaborated including validity issues and ways to overcome the gaps in the behavioral literatures that confront developers of integrated cognitive models.1. IntroductionA common concern amongst agent developers is to increase the realism of the
7、 agents behavior and cognition. In training, wargaming, and operations rehearsal simulators there is a growing realization that greater cognitive subtlety and behavioral sensitivity in the agents leads to both (1) a greater ability to explore alternative strategies and tactics when playing against t
8、hem and (2) higher levels of skill attainment for the human trainees: e.g., see 1 and 2. For this to happen, the tactics, performance, and behavior of agents must change as one alters an array of behavioral and cognitive variables. For example, one would like agent behavior to realistically change a
9、s a function of: the culture they come from (vital for mission rehearsal against forces from different countries); their level of fatigue and stress over time and in different situations; and/or the group effectiveness in, say, the loss of an opposing forces leader. At present, however, this does no
10、t happen, and in most of the available combat simulators the agents conduct operations endlessly without tiring, never make mistakes of judgment, and uniformly (and predictably) carry out the doctrines of symmetric, sometimes vanquished opponents, such as the Warsaw Pact, among others.Closely relate
11、d to the topic of emulating human behavior is that of “believability of agents. The basic premise is that characters should appear to be alive, to think broadly, to react emotionally and with personality to appropriate circumstances. There is a growing graphics and animated agent literature on the b
12、elievability topic (e.g., see 3, 4 and 5), and much of this work focuses on using great personality to mask the lack of deeper reasoning ability. However, in this paper we are less interested in the kinesthetics, media and broadly appealing personalities, than we are in the planning, judging, and ch
13、oosing types of behavior - the reacting and deliberating that goes on “under the hood of embodied agents. Finally, and perhaps most importantly the human behavior literature is fragmented and it is difficult for agent developers to find and integrate published models of deeper behavior. Our research
14、 involves developing an integrative framework for emulating human behavior in order to make use of published behavioral results to construct agent models. We are not attempting basic research on how humans think but on how well existing models might work together in agent settings. That is, the fram
15、ework presented here is intended for experiments on how to integrate and best exploit published behavioral models, so as to improve the realism of agent behaviors when one seeks to model individual differences such as stress, emotion, and culture.In particular, we are interested in emergent macro-be
16、havior due to micro-decisions of bounded-rational agents and with developing a framework that promotes the study of specific phenomena (i.e., emotions, stress, and cultural values) that lead to limits of rationality. What motivates agents to select actions that sub-optimize their own utility as well
17、 as that of groups whose causes they seek to advance? To explore this question, we have been researching a framework that allows one to investigate the duality of mind-body interaction in terms of the impact of environment and physiology on stress and, in turn, of stressors on rationality. Our frame
18、work also attempts to integrate value systems and emotion-based appraisals of decision options along with the stress constraints. That is, we have been working towards a framework that permits one to examine the impacts of stress, culture, and emotion upon decision-making. With such a framework, one
19、 should, for example, be able to readily model and visually render what makes one protesting crowd throw stones while another peacefully demonstrates.As soon as one opens the door to modeling the impact of stress, culture, and emotion on rationality, one must be amenable to the idea that competing v
20、iews, results, models, and approaches have to be examined and potentially integrated. The point of such a research program should not be to argue for one approach or theory over another, but to provide ways to readily study alternative models of whatever contributes to the phenomena of interest. 1.1
21、 Role of Emotion and Cultural Value Ontologies in Agent Behavior“Emotive computing is often taken to mean the linking of the agent state to facial and body expressions, vocal intonation, and humorous or quirky animation effects: e.g., see 6, 7 and 4. However, recent theories identify emotions as vit
22、al to the decision-making process and to manage competing motivations 8. According to these theories, integrating emotion models into our agents will yield not only more believable decision-makers, but also more realistic behavior by providing a deep model of utility. These agents will delicately ba
23、lance, for example, threat elimination versus self-preservation in much the same way it is believed that people do. These theories suggest that without adding emotional construal of events, the agents wont know what to focus upon and what to ignore, and wont know how to balance the set of next-step
24、alternative actions against larger concerns, as in the case of Damasios patients with pre-frontal cortex damage who spend the entire day mired in highly logical decision analyses of banalities, even at the cost of their own self-interest and survival.Important implementations of these ideas and theo
25、ries were attempted in the “believable agents movement such as 4 and 5 which seek to improve the believability of characters behavior in fictional settings with the help of an emotion model. The OCC model is probably the most widely implemented of the emotion models (e.g., 9, 10 and 11) and it expla
26、ins the mechanisms by which events, actions, and objects in the world around us activate emotional construals. In both Oz 4 and the Affective Reasoner 5 projects, emotion was largely modeled as a reactive capability that helped characters to recognize situations and to reflect broad and believable p
27、ersonality characteristics. Later versions of Oz include a behavior planner, but the link between emotion construals and behavioral choice is never well articulated in their published accounts. On the other hand, 12 and 13 concretely extend the OCC model via the use of an event planner into a deeper
28、, deliberative reasoning mode where agents were able to construe the value of plans and plan elements (events that havent happened yet). In the current paper we ignore multi-step planning, although we leave open the possibility of re-introducing it in a future revision of the architecture, but as a
29、result our agents are able to construe the impact of objects and behavior standards both on themselves and on those they like/dislike. We go beyond this too to the area of what is probably unconscious construals of stressors such as fatigue, time pressure, and physiological pressures. This means we
30、attempt a reasonably full implementation of the OCC model for reactions and deliberations of all types of events, actions, and objects. This approach provides a generalizable solution to another issue in the OCC model. The OCC model indicates what emotions arise when events, actions, or objects in t
31、he world are construed, but not what causes those emotions or what actions an agent is likely to take as a result. There is no connection between emotion and world values (cultures), even though other theories suggest such a link 8, 10 and 11. In contrast, cultural concern or value ontologies are re
32、adily available in the open literature (e.g., the ten commandments or the Koran for a moral code, military doctrine for action guidance, etc.) and may readily be utilized to implement an agent of a given type in the framework we present here. Ideally, one would like to tie such cultural concern onto
33、logies indirectly to the emotional processes of the agent, so that situation recognition as well as utilities for next actions are derived from emotions about ontologies and so that both reacting and deliberating (judging, planning, choosing, etc.) are affected by emotion. 2. Cognitive Architecture
34、and Framework The research described here is not to propose the best cognitive architecture or agent algorithms but to propose a reasonable framework within which the many contributions from the literature can be integrated, investigated, and extended as needed. There are a large number of similar f
35、rameworks in the literature. A useful comparison of 60 such models may be found in Crumley & Sherman 14.Our framework is built around a blackboard data structure loosely corresponding to a short-term or working memory system. Modular PMF subsystems manipulate data contained both in the blackboard an
36、d in a long-term memory store. Information is layered on the blackboard such that each layer is dependent on the layers below it for a given decision cycle of the agent (see Figure 1).Figure 1 Overview of the Integrative Architecture for Researching Alternative Human Behavior Models for Generic Agen
37、tsOne of the principal strengths of our system is this modularity. In and of itself, PMFServ lacks a number of notable functions that are not necessary for our current explorations into the emergent properties of social agents, but might be essential for future applications of the system. For exampl
38、e, our agents themselves do not plan future actions or anticipate what other agents might do in upcoming time steps. Also, our agents are not designed to model expert behavior on any specific task. There are existing systems that accomplish these and other exclusively cognitive tasks very well (e.g.
39、 SOAR15, ACT-R16, etc). Rather than compete with such systems, PMFServ can interact with them, using its physiology, perception, and emotion PMF subsystems to constrain the decisions generated by an alternate cognitive architecture. We are also considering developing alternative PMF subsystems that
40、would provide decisions across a wide spectrum of realism and depth. For example, our current set of PMF modules would provide too much detail and be far too processor-intensive if we were trying to simulate very large crowds of agents. To this end, we have been developing a set of PMFs based on cel
41、lular automata (see 17, 18 for some background on this approach) that allow us to simulate massive crowds of simple agents that can still interact with a smaller set of more complex agents.The following is a description of our architecture with PMF modules designed to support a modest number (at mos
42、t 100) of social agents. Our description will follow the decision cycle of a single agent as it progresses up the blackboard.2.1 Physiological SubsystemsThe first and most fundamental input to the agent is its own physiology. We model a series of physiological reservoirs that maintain the level of s
43、tress the agent is experiencing as a result of its physical environment and the state of its body. The reservoirs currently in use are: exertion, injury, temperature, sleep, and nourishment. Each is grounded in a valid PMF. Presumably, the agents physical state will have changed somewhat between dec
44、ision cycles and this physiological change is reflected at the very beginning of the new cycle. If the agent is dead, unconscious, or in shock, this will clearly have an affect on its cognitive ability. Less extreme physical states are important as well, however, as they help to determine the agents
45、 overall arousal.The agents physiology is based around an energy reservoir, or tank. We have developed an energy consumption model based on the SAFTE and FAST models described by Douglas and Hursh19. As the agents desired arousal and magnitude of physical exertion change, the agent opens and closes
46、a valve at the bottom of the tank that releases the energy to be used for those tanks. The agent is bound by the flow of energy out of the tank. For example, if the supply of energy in the tank is quite low, the flow out of the tank may not be sufficient to support a high-exertion task. This system
47、gives us a way to moderate performance on tasks as well. If an agent wants to sprint consuming energy at a burst rate of 950 kCal/hr but does not have sufficient energy to produce a 950 kCal/hr flow from its tank, we can use the flow to determine how fast the agent can actually run as it gets increa
48、singly tired. We used data from US Army Field Manuals20 to establish the relationship between physical load, energy consumption, and fatigue, and used that relationship to calibrate our model. Figure 2: Physiology ModuleEach physiology PMF ties into this system. A virtual stomach refills the energy
49、tank based on the agents rate of digestion. When the agents sleep falls below a critical threshold, a second valve in the tank of energy is opened. As the agent becomes increasingly sleep-deprived the valve opens further, resulting in gradually increased inefficiency and decreased performance on all tasks 19. Injury and temperature may both affect the