COMPUTER ANIMATION AND VIRTUAL WORLDS Comp. Anim. Virtual Worlds (2009) Published online in Wiley InterScience (www.interscience.wiley.com) DOI: 10.1002/cav.316 ******************************************************************************************************************* Development of a computational cognitive architecture for intelligent virtual character By Pak-San Liew*, Ching-Ling Chin and Zhiyong Huang ********************************************************************************************* A development of a computational cognitive architecture for the simulation of intelligent virtual characters is described in this paper. By specializing and adapting from an existing structure for a situated design agent, we propose three process models—reflexive, reactive and reflective—which derive behavioural models that underlie intelligent behaviours for these characters. Various combinations of these process models allow intelligent virtual characters to reason in a reflexive, reactive and/or reflective manner according to the retrieval, modification and reconstruction of their memory contents. This paper offers an infrastructure for combining simple reasoning models, found in crowd simulations, and highly deliberative processing models or reasoning, found in ‘heavy’ agents with high-level cognitive abilities. Intelligent virtual characters simulated via this adapted architecture can exhibit system level intelligence across a broad range of relevant tasks. To demonstrate the usefulness of the proposed architecture, we describe the effect of the reflexive, reactive and reflective processes on a virtual character in our virtual tour application. Copyright # 2009 John Wiley & Sons, Ltd. Received: 25 March 2009; Accepted: 26 March 2009 KEY WORDS: computational cognitive architecture; virtual exploration; intelligent virtual character; simulation Introduction Simulating a believable character in virtual world is one of the major research areas in computer graphics and artificial intelligence. There are many applications such as computer games and virtual reality systems, where a believable character can greatly enhance the immersive experience of a user. The current use of artificial intelligence technology to produce intelligent virtual characters has yielded interesting results. However, most of the existing work produced characters that exhibit a narrow spectrum of intelligence. This has led researchers to consider a new, holistic and humaninspired approach to inject intelligence into virtual characters. In this work, we develop an infrastructure for intelligent virtual characters to perform a potentially broad range of tasks—a highly desirable attribute for *Correspondence to: P. S. Liew, Institute for Infocomm Research, 1 Fusionopolis Way, 21-01 Connexis, South Tower, Singapore 138632, Singapore. E-mail: [email protected] simulating believability in virtual characters within interactive digital media. Based on an architecture described by Liew and Gero,1,2 we propose three process models—reflexive, reactive and reflective—which derive behavioural models that underlie intelligent behaviours for these characters. Quick stimulus responses are considered as ‘reflexive’; stimulusresponse behaviours with slight deliberation are considered as‘reactive’ while highly deliberative responses are considered as ‘reflective’. Simulating reasonings in the reflexive, reactive and reflective manners are made possible by combinations of these process models, according to the retrieval, modification and reconstruction of their memory contents. Our adapted architecture provides the capability of combining simple reasoning models, found in crowd simulations, with highly deliberative processing models, found in ‘heavy’ agents with high-level cognitive abilities, within a unified framework, hence extending the range of intelligent behaviours that can be exhibited in virtual characters. An initial implementation of a virtual tour application— TOWN, based on a behavioural model derived from ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. P. S. LIEW, C. L. CHIN AND Z. HUANG ******************************************************************************************************************* this architecture is demonstrated. This is the first step towards representing and processing an agent’s knowledge and experience to simulate the required humanlevel intelligence for believable virtual characters. Related Work This research approaches the development of intelligent systems by using humans as the best example of humanlevel intelligence through the process of computational cognitive modelling. Computational cognitive modelling looks at the essence of cognition and various cognitive functionalities by specifying computational cognitive models.3 Research in cognitive modelling may revolve around modelling human cognition to be tested empirically with human subjects, or it may use the result as an effective mechanism to build intelligent systems.4 The building of intelligent systems is the focus of this paper. A computation cognitive model captures internal computational processes that generate overt cognitive behaviour.5 Cognitive models can be created with or without a cognitive architecture. A cognitive architecture specifies the underlying infrastructure for an intelligent virtual character that is constant over time and across different application domains.4 The importance of a cognitive architecture arises from its provision of a comprehensive initial framework for further development of cognitive models in many task domains. As a result of initial assumptions made in designing the architecture, further development of models is constrained by the pre-determined modular structure within unified framework to produce general intelligence.6 There are research works that develop intelligent systems independent of a cognitive architecture.7–11 However, this paper takes the position of the development of an architecture of intelligent behaviour through a system-centric approach. This architecture will have the potential to unify various findings in artificial intelligence, such as production rules in expert systems, and cognitive science, such as information processing models of the human mind, into a single framework. We aim to fulfil the increasing requirements of integrated systems that wish to support the required behaviour across a broad range of relevant tasks.4 An existing structure for a situated design agent1,2 is specialized and its related key concepts are adapted for simulating believable virtual characters. Development of a Computational Cognitive Architecture The approach taken in this paper, to develop the required computational cognitive architecture for an intelligent virtual character, is based on creating those aspects of the character that are constant over time and across different application domains.4 These include a memory system that contains knowledge and experience processed by the character, representations of elements that are contained in this memory system and their organization into largerscale mental structures and functional processes operating on these structures, performance mechanisms utilizing them and learning mechanisms that change them. An existing structure1,2 for a situated design agent (Figure 1) is specialized and related key concepts are adapted for developing the computational cognitive architecture in this paper. This new architecture is illustrated in Figure 2. Various components of the intelligent virtual character are represented as boxes in the diagram, with paths for information flow labelled with numbers. Processing within the architecture is organized around different computational cognitive systems labelled as conceptor, perceptor, sensor and memory system in Figure 2. Each of these systems contributes Figure 1. An existing structure for a design agent used for the development of the cognitive architecture in this paper (from Reference1,2). ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav A COGNITIVE ARCHITECTURE FOR VIRTUAL CHARACTER ******************************************************************************************************************* Figure 2. Computational cognitive architecture developed in this paper for intelligent virtual character (numbered paths are indicated as (#) in the text where # is the number as shown in the diagram) (expanded and specialized from Reference1,2). towards the operations of reflexive, reactive and reflective process models to create behavioural models (Figure 3). Details of these cognitive systems, process models and behavioural models are described in the following sub-sections. Computational Cognitive Systems: Sensor, Perceptor and Conceptor Sensor, perceptor and conceptor1,2 are different subsystems within the architecture that provide functionalities for sensation, perception and conception. Sensation is an information extraction process. Information from the environment is extracted into the agent according to the current configuration of the sensor. This configuration is dictated by perception that biases the way the sensor works to limit the type and amount of information sensed. Perception, after creating the bias for sensation, organizes the sensory data into coherent structures as required by conception. It operates on existing knowledge and experiences to form concepts relevant for the current situation. Each of these processes, sensation, perception and conception, draws upon information from the memory system to produce actions through the effector. Memory System Figure 3. Creation of different behavioural models from the cognitive architecture by various configurations of process models. The purpose of the memory system is to provide an environment for the storage, retrieval, modification and ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav P. S. LIEW, C. L. CHIN AND Z. HUANG ******************************************************************************************************************* construction of memory out of knowledge and experience1. The right-hand side of Figure 2 illustrates details of the memory system expanded from Liew and Gero’s1,2 work on situated design agents. A brief description of each memory sub-system follows. Working memory is a workspace for deliberative processes, where explicit reasoning and executive control functions are performed with caching (11)2. These processes selectively alter and/or add information to the working memory based on the content of long-term memory (LTM) (14) or short-term memory (STM) (12). Information, within the working memory, is combined with the stored knowledge and experiences, manipulated, interpreted and recombined to develop new knowledge, assist learning, form goals and support interactions with the external environment. Knowledge and experiences accumulated are available for use in the immediate task and once activated, they can then be manipulated for the extraction of novel information that had not been stored explicitly in the past. The behaviours of these operations are influenced by the external environment, the interactions between agent and environment, and the internal state of the agent. All the above inputs are subjected to biases (13) before they are processed within the working memory system. The bi-directional path (13) between the priming component and the working memory models the effect of using the default bias or the setting of a pre-defined bias. Short-Term Memory is modelled as a short-term, fixedsize buffer that captures the limited-capacity characteristics of STM. The content of the buffer is constructed from the memory trace that enters the STM (16). The path (17) from the buffer to the LTM indicates the buildup of a long-term trace due to the memory trace’s levelof-processing. The slots in STM are used to indicate the notion of the relative level-of-processing of the various traces within the buffer. 1 Knowledge refers to the general facts related to that domain that are given to the agent when it was first created. The agent is not involved in any way with these facts. Experience of an agent entails the agent’s involvement as the ‘first person’ in dealing with the substance of that experience. Memory is a ‘thing’ that involves a construction process for its formation. Memory construction is the process that produces a memory by manipulating knowledge and experience. It requires a total reorganization of what the system knows. Memory retrieval is a process that treats the memory system as a static store, where required knowledge is retrieved according to simple matching between the stimulus from the environment and the preconditions stored within the memory system. Memory modification is a process that entails minor adaptation of retrieved knowledge for the current situation. 2 Numbers in parentheses refer to the paths in Figure 2. Both STM and LTM are used in the construction of memories within the working memory in the reflective and reactive mode of operations ((12) and (14)). The differences between memory traces residing in STM and LTM lie in the fact that memories in STM are not as readily recalled in the memory construction process as those in LTM. The STM serves as a gateway into and out of the LTM. The path into LTM occurs when memory traces are transferred from the working memory to the LTM through STM. As soon as a memory of an experience has been constructed and served its purpose within the working memory, it is transferred into STM as a new memory trace (12). Memory traces in STM are integrated into LTM gradually through their level-of-processing (17). The STM acting as a gateway out of the LTM occurs when memory traces in LTM that are not grounded frequently are transferred into STM (16) and out of the memory system. These two actions model the phenomenon of long-term learning and the phenomenon of ‘use it or lose it’ in the human memory system. The amount and form of memory traces that are transferred from STM to LTM (17) are primarily a function of their level-of-processing. For a memory trace to be transferred to LTM, it must go through a series of encoding levels. Active processing of a memory trace through its grounding, according to the current situation, produces durable STMs. As the durability of a memory trace increases, its level of encoding increases. The increase in level-of-processing of a specific memory trace facilitates its inclusion into LTM. This simple formulation captures the important characteristic of long-term learning in which a process of encoding material, in terms of prior experiences and grounded usage, produces comparatively durable and readily retrievable memory traces. Long-Term Memory is a relatively permanent repository for memory traces. Memory traces stored here do not disappear as readily as those in STM. New memory traces about experiences are encoded into LTM after they have been processed in the working memory and STM. To influence subsequent explicit behaviour, these memory traces within LTM and STM are brought back into working memory, where they become part of the reasoning process. Long-term memory refers to all effects of prior experiences and knowledge on subsequent behaviours of the situated agent. These behaviours are instantiated through the various reflective, reactive or reflexive processing models (Figure 3). Explicit memory subsystems provide memories for reasoning involving the ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav A COGNITIVE ARCHITECTURE FOR VIRTUAL CHARACTER ******************************************************************************************************************* working memory (for reflective and reactive process models) while implicit memory sub-systems deal with memories that affect the reflexive behaviour of the agent (reflexive process models). Explicit memory provides the required information for reflective and reactive processing models. Requests for explicit memories and transfer of constructed memories occur through process (14). These operations are subjected to the bias activated at the current point in time (13). Explicit memories that are not subsequently grounded are transferred to STM before they are purged from system (16). On the other hand, memories in STM that have passed through a series of processing (levelsof-processing) are transferred to explicit memory (17) for longer retention. Episodic memory deals with the storage and construction of memory traces related to experiences. Once the memory of an experience has been constructed and used within the working memory, it is stored in STM as a new experience where it is subjected to several levels of processing before it is transferred into the episodic memory (17). The content of the episodic memory is also subjected to generalization (20) where a common set of experiences is compiled into a general rule and the original experiences may be removed from the system. Semantic memory deals with general principles and refers to the acquisition, storage and usage of knowledge. This knowledge is grounded according to the agent’s current internal state in relation to the current environment and is subjected to addition, modification, updating, replacement and removal as the agent learns new knowledge. Implicit memory captures knowledge expressed through performance rather than through recall and adaptation (semantic memory) or construction (episodic memory). Experiences are accumulated in behavioural changes. The reflexive behaviour of the agent is based on the implicit memory system. Implicit memories are taken as memories that are not used explicitly within the working memory during reasoning. These memories are previously acquired information that influences the agent’s behaviour or reasoning process either directly or indirectly. Implicit memory involves a whole range of implicit learning systems that can act as the basis for analysing perceptual and motor processing (10), and they typically involve relatively automatic retrieval processes that are normally not carried out during reflective and reactive processes. Feedbacks from the sensor (10) form the basis of these implicit learning processes. With practice, implicit skills can be improved. The improvement is retained (remembered) from one practice session to the next even though the information remembered occurs outside the working memory. In other cases, a pattern of behaviour that was learned through the working memory can gradually become automatic or habitual with repetition, causing memories that were once explicit to become implicit. Thus implicit memory can be used to carry out tasks in a reflexive manner, leaving the reflective and reactive processes free for other tasks (18). The contents of implicit memory are also subjected to modification by the working memory through a series of explicit learning processes. These processes add, remove and alter the contents of implicit memory through normal machine learning processes (18). Implicit memory has a constraining characteristic of extreme dependence on the immediate situation. A specific behaviour is automatically called upon by a specific configuration of stimuli. The same behaviour is not involved in another set of stimuli. Explicit memories, on the other hand, are used in contexts other than those in which they were originally acquired and are therefore used not only to guide repetitive behaviour but also to plan future behaviour and to modify existing routines. Procedural memory contains process knowledge on how to perform a specific task. This knowledge about processes can be taken as the skill that an agent possesses. With practice, these skills can be improved. The improvement is retained (remembered) from one practice session to the next using feedback from the working memory (18), even though the information remembered is used in a reflexive manner later ((3) via (10)). If necessary, the knowledge learned is used together with that in the semantic memory if its application into the current situation needs to be considered together with other knowledge. Priming is the process that activates (biases) one or more existing memory traces by a stimulus (the priming stimulus) presented to the working memory system (13), and it influences the subsequent perception (9) and reasoning processes (8). The priming stimulus originates from memory traces contained within LTM or STM ((15) and (19)) and it allows activated memories in STM or LTM to be brought into working memory more easily after the priming experience than before ((12) and (14)). Priming plays an important role in the explicit reasoning process. When an object, idea (concept) or event is seen (through perception) or thought about (within the reflection and reactive processes), those elements of the semantic memory, that are relevant to that particular perception or thought, are activated ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav P. S. LIEW, C. L. CHIN AND Z. HUANG ******************************************************************************************************************* (primed) for a period of time and made more available than usual for retrieval to the working memory for deliberation. This makes the reflective and reactive processes operate logically rather than jumping randomly from one idea or image to another. Perceptual (form) priming ((9) and (13)) can improve the facility for detecting or processing perceptual objects based on recent experience. It can involve the acquisition of new information or the modification of existing memory representation that not only improves the ability to identify stimuli but also alters judgments and preferences that are involved in the same stimuli. Conceptual (meaning) priming ((8) and (13)) affects the agent’s behaviour indirectly by influencing the reflection and reaction processes. It influences the processes that are carried out within working memory and involves the activation of concepts stored in the semantic memory system. Conditioning provides reflexive behaviour to the agent by matching the pre-conditions of the rules it contains with the conditions of the environment and applying those rules when applicable. The agent’s behaviour is affected directly by this reflexive action in the reflexive operation. Common reactive operations that are frequently used are transferred to the conditioning component (18) so that they can behave in a reflexive manner without explicit reasoning in the working memory. The if-then rules, in the conditioning component, are extremely dependent on the immediate situation. An implicit memory is retrieved and used to guide behaviour only at the moment when a specific set of pre-conditions is satisfied. The learning of new rules either by creating entirely new rules or by modifying/replacing existing rules will result in the creation of an artificial grammar. This occurs through an implicit learning process such as classical conditioning or operant conditioning. Effector and Environment The effector output actions that are performed within the environment to produce effects that can be sensed (left-hand side of Figure 2). Structural design of this system in relation to the cognitive architecture is based on Blumberg and Galyean’s12 layered architecture for behaviour, motor system and geometry direction of autonomous creatures and Reynolds’s13 action selection-steering-locomotion hierarchy of motion behaviours. The effector operates at two levels: a controller layer and an actuator layer. At the controller level, motion commands resulting from the reflective, reactive and/or reflexive processing models are integrated and converted to control signals to be sent to the actuator for execution of actions (21). The actuator is modelled after Reynolds’s13 locomotion level. It represents the agent’s embodiment within the simulated world. Within the actuator, control signals from the controller are converted into motions according to the agent’s embodiment. The two-layered design of the effector allows different actuators to be used without affecting the rest of the architecture. If an appropriate convention for communicating control signals has been established, the entire architecture can be completely independent of the specific actuator of similar functionalities. When a new actuator is used, only the mechanism for mapping control signals to motions that is unique to the agent’s new embodiment needs to be altered. Actions performed by the actuator are subjected to simulated physical constraints imposed by the environment on the agent’s embodiment. In order to compensate this interaction between the agent’s body and the environment, the control commands can be tuned according to the actuator used through self-calibration. Knowledge of such self-calibration skills can be called upon reactively (9) or reflectively (8). Reflexive, Reactive and Reflective Process Models There are three basic process models that can be derived from the cognitive architecture of agency: reflexive, reactive and reflective process models. These are extensions of processes of the same name the operate within virtual agents.14 Reflexive process model represents processes that are very fast and operate in a parallel fashion. They are either innate abilities of the agent or are learned environmental competencies and behaviours. Operations of reflexive processes are based on a direct response to stimulus sensed by the agent (7, 10, 3, 1)3. Sensory data about the environment and about the internal state of the agent are acquired through the sensor and propagate without perception to effectors. A reflexive process cannot devise, evaluate and choose alternative possible actions in advance of performing 3 Sequence of numbers in parentheses refers to the information flow paths in Figure 2. ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav A COGNITIVE ARCHITECTURE FOR VIRTUAL CHARACTER ******************************************************************************************************************* them. Various forms of implicit learning/training can extend the capabilities of the reflexive sub-mechanism. A major characteristic of the reflexive process is that all responses, internal within the agent or external to the environment, are activated as soon as their triggering conditions are met. This process can break down if the agent is presented with a condition that was not predefined for the system. The main purpose of reactive processes is to tune parameters for correct behaviours by devising new combinations of actions to cope with novel contexts. During the operation of a reactive process within the working memory, a model is created to evaluate alternative actions before they are carried out (7, 6, 9, 4, 1). A number of different temporary structures such as alternative plans may be created and compared in some way prior to selection. The operation of a reactive process does not involve additional processes that create new functional capabilities; rather it uses a general sub-system to create and evaluate different configurations of the current capabilities. Operations in the reactive process that are called upon frequently can be made to behave in a reflexive mode (bypassing the reasoning process) and run in parallel with other reactive process. The concurrency of these operations allows routine tasks to be carried out automatically, leaving the reasoning process to deal with tasks that are more complicated and thus facilitating better overall performance of the system. Reflective process model represents processes that interact with other internal processes, within the working memory, to provide an overall behaviour of the agent (7, 6, 5, 8, 2, 1) by behavioural parameter reparameterization and expansion. Agents operating in the reflective mode monitor, evaluate and modify other processes occurring within the agent to regulate and direct the internal and external behaviour of the agent. The reflective processes within the agent give it the abilities to reason about past events, make predictions about future events, form concepts, and act in a proactive manner according to some goals. that span from instant stimulus-response (found in crowd simulations) to high deliberation (found in ‘heavy’ agents with high-level cognitive abilities). To illustrate the concept of the behavioural model, we demonstrate the change in behaviour of a virtual character under the influence of reflective, reactive and reflexive processes in an in-house application, named ‘TOWN’, in the coming sections. The next section also draws the parallel between the behavioural model concepts and the algorithm that drive the virtual character. Initial Implementation TOWN (pronounced dot-town) is a virtual town exploration application. It introduces Singapore landmarks, representing them in this virtual world by caricatured models with low polygon count. A virtual character, a bear, roams around in the virtual world and brings its audience to the different places of interest. Interaction from users is minimal—users can stop/run the bear, look at the surrounding of the bear using different camera views and ‘teleport’ the bear to another location. In the application, steering behaviours13,15 are used to control the movement of the bear. Movements such as seek, arrive, obstacle avoidance and wander are modelled as forces acting on the bear that influence its movement. These forces are scaled (Figure 4) according to their importance before combining as a resultant force that steer the bear in a particular direction with a specific velocity. Figure 4 illustrates an example of four steering forces (F1, F2, F3 and F4) calculated from a behavioural model acting on an agent. Each of these forces is scaled by a factor (a, b, g, d) that represents its influence on the motion of the agent. The behavioural model used here Behavioural Model Behavioural models represent the lowest level of abstraction in the architecture. These models are used directly to control a virtual character’s behaviour. A behavioural model can be created out of a combination of different reflective, reactive and reflexive processes (Figure 3). This gives the character reasoning capabilities Figure 4. Steering forces scaled by their influence on the motion of the agent. ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav P. S. LIEW, C. L. CHIN AND Z. HUANG ******************************************************************************************************************* utilizes the reflective, reactive and reactive process models. At the fundamental operational level, all steering forces are calculated by algorithms stored within the procedural memory according to sensory data extracted by the sensor from the virtual environment (italic text refers to different modules illustrated in Figure 2). These forces are sent to the controller (10, 3) as motion commands together with default scaling from the sensor and sum together in the default order as determined by the controller. A control signal modelled as a resultant force acting on the bear is sent to the actuator (21) for the bear to execute its movement. This process utilizes simply the effector, sensor and procedural memory (7, 10, 3, 1) in a reflexive manner. There are two fundamental ways to control the movement of the agent through steering forces. The first is by changing the influence of each steering force (Figure 5). The second is by adding new steering forces (Figure 6). Changing the influence of each steering force can be seen as correcting errors in movement such as the bear crashing into walls or the bear bouncing between two parallel obstacles. The behavioural model operates in a reactive mood when this happens (7, 6, 9, 4, 21, 1). The perceptor can be set to look out for signs of error movement by interpreting the sensor data (6) and trigger an event to extract relevant knowledge from semantic memory (9, 14) to modify the influence of each steering force coming from Procedural Memory (19, 14, 9). Forces that are subsequently sent to the controller (4) as motion commands are modified with new scaling from perceptor and summed in an order specified by the controller to create the necessary controller signal for the actuator (21) to correct the movement of the bear. Figure 6. Adding new steering forces to control the agent’s movement. Adding new steering force to control the bear’s movement can model achieving of high-level goals through reflective reasoning (7, 6, 5, 8, 2, 21, 1). In this reflective mode, the conceptor can create high-level goals by processing the information from the perceptor according its knowledge stored in semantic memory (6, 5, 8, 14). For example, the bear may have an energy level attribute (computed from selected sensory data based on knowledge) that starts to deplete as the bear moves in the environment. As soon as a certain level has been reached, a ‘hunger’ emotion may be triggered to create a ‘find food’ goal. This goal can create additional forces in the algorithms within procedural memory (19) that steer the bear towards specific location in order to find food. When the time comes, all of these forces are sent to the controller (19, 14, 8, 2) as motion commands together with the required scaling from the conceptor and summed together in the order as determined by the controller. A control signal modelling the resultant force that pulls the bear towards the landmark with food is sent to the actuator (21) for the bear to execute its movement. Results Figure 5. Changing the influences of different existing steering forces to control the agent’s movement. Up to this point, much has been said of the reflexive, reactive and reflective processes. In our visual simulation using the TOWN application, the influence of these three processes on the bear is demonstrated. The effect of reactive adjustment is best seen when the bear is in a confined space. We placed the bear in two locations in the virtual world (Figures Figure 7(a) and 7(b)) where its running space is small. Without reactive adjustment (and hence in reflexive mode), the bear is easily trapped ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav A COGNITIVE ARCHITECTURE FOR VIRTUAL CHARACTER ******************************************************************************************************************* Figure 7. (a) (b) The confined spaces in the virtual world. in the confined area, bouncing from one wall to another. In the reactive processing, the scales of the steering forces are adjusted, correcting errors so that the bear can manoeuvre itself out of the confined space. The visual simulation also demonstrates the influence of reflective process, which is adjusted by creating additional forces that will steer the bear towards a designated destination. In the visual simulation, the bear wanders aimlessly when there is no reflective adjustment. When the reflective process is turned on, the bear moves towards its target, manoeuvring itself out of any obstacles that come in its path. We see from the visual simulation that with the adjustment of the reactive and reflective processes, the virtual character shows intermediate and goal-directed form of reasoning. The concept of the behavioural model of the proposed architecture enables the bear to exhibit more human-like behaviour, hence increasing the appeal and usefulness of the application. Conclusion and Future Work In this paper, our development of a computation cognitive architecture for intelligent virtual characters has been described. To demonstrate the effectiveness of the development, a behavioural model that utilizes the reflexive, reactive and reflective processes was used to modify a common steering mechanism in our virtual tour application. The proposed architecture has the potential to simulate human-like behaviours in intelligent virtual characters. Future work will be on the representations for items stored in each memory sub-system, and processes that operate on these representations of the proposed architecture. More realistic human-like behaviours can be simulated in this way. ACKNOWLEDGEMENTS The work is supported under the Personal 3D Entertainment Systems Programme of the Institute for Infocomm Research, ASTAR. The TOWN came into existence with contributions from Bryan Chong, Chua Gim Guan, Loke Mei Hwan, Christina Tang, Camellia Zakaria, Azmi Karim and Loh Jin Biao.The authors are very grateful to these colleagues. References 1. Liew P-S, Gero JS. A memory system for a situated design agent based on constructive memory. In Proceedings of CAADRIA2002, Eshaq A, Khong C, Neo K, Neo M, Ahmad S (eds). Prentice Hall, New York, 2002; 199–206. 2. Liew P-S. A constructive memory system for situated design agent. PhD Thesis, Key Centre of Design Computing and Cognition, University of Sydney, 2004. 3. Sun R. Introduction to computational cognitive modeling. In The Cambridge Handbook of Computational Psychology, Sun R (ed.). Cambridge University Press: NY, 2008; 3–19. 4. Langley P, Laird JE, Rogers S. Cognitive architectures: research issues and challenges. Cognitive Systems Research 2009; 10(2): 141–160. 5. Sun R, Ling CX. Computational cognitive modeling, the source of power, and other related issues. AI Magazine 1998; 19(2): 113–120. 6. Sun R. Desiderata for cognitive architectures. Philosophical Psychology 2004; 17(3): 341–373. 7. Shao W, Terzopoulos D. Autonomous pedestrians. Proceedings of SIGGRAPH/EG Symposium on Computer Animation (SCA’05), 2005; 19–28. 8. Funge J, Tu X, Terzopoulos D. Cognitive modeling: knowledge, reasoning and planning for intelligent characters. Proceedings of SIGGRAPH 99, 1999; 29–38. 9. Funge J. AI for Games and Animation: A Cognitive Modeling Approach. A K Peters: Natick, MA, 1999. 10. Park A, Calvert T. A social agent pedestrian model. Computer Animation and Virtual Worlds 2008; 19: 331–340. 11. Luo L, Zhou S, Cai W, et al. Agent-based human behavior modeling for crowd simulation. Computer Animation and Virtual Worlds 2007; 19: 271–281. ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav P. S. LIEW, C. L. CHIN AND Z. HUANG ******************************************************************************************************************* 12. Blumberg BM, Galyean TA. Multi-level direction of autonomous creatures for real-time virtual environments. Proceedings of 22nd Annual Conference on Computer Graphics and Interactive Techniques, 1995; 47–54. 13. Reynolds CW. Steering behaviors for autonomous characters. Proceedings of Game Developers Conference 1999, 1999; 763–782. 14. Mahar ML, Gero JS. Agent models of 3d virtual worlds. Proceedings of ACADIA 2002: Thresholds, 2002; 127–138. 15. Buckland M. Programming Game AI by Example. Wordware: Plano, TX, 2005. Authors’ biographies: Pak-San Liew received his PhD in Design Computation from the Key Centre of Design Computing and Cognition, University of Sydney in 2004. He is currently a Senior Research Fellow in Institute for Infocomm Research, ASTAR. His research interest is in artificial intelligence and computer graphics. Chin-Ling Chin received her Masters in Electrical and Electronic Engineering from Imperial College, UK in 1997. She worked in DSO National Laboratories, Singapore, in speech processing before joining Institute for Infocomm Research, ASTAR, as a Senior Research Officer. Her research interest is in computer graphics. Zhiyong Huang received his PhD in Computer Science from EPFL, Switzerland in 1997. He worked in School of Computing, NUS, Singapore, before joining Institute for Infocomm Research, ASTAR, as a Senior Scientist. His research interest is in computer graphics. ******************************************************************************************************************* Copyright # 2009 John Wiley & Sons, Ltd. Comp. Anim. Virtual Worlds (2009) DOI: 10.1002/cav
© Copyright 2026 Paperzz