|
||||
|
A cognitively inspired model for self-aware agents
| Permalink The human brain arose as a result of natural and sexual selection operating in the typical human framework of social systems. One of the main tasks humans have had to tackle for evolutionary survival is that of self-awareness, essentially answering questions such as: “Who am I?” “Where am I?” “What context am I in?” and “What are other people doing, and why?” Psychologists, sociologists and neuroscientists have extensively studied the typical behaviour of humans in a variety of situations, and distilled a large body of knowledge from such research. However, because these findings are rarely expressed in mathematical or procedural form, they require further work before they can be transferred to an information technology (IT) context.1, 2 We know (mainly from observations) what the characteristic human responses to stimuli are and how people generally make decisions. It is also well established that in many cases the human response is not rational, because it relies on efficient but simplified experiental rules, or ‘heuristics’, rather than on exhaustive analysis. Such a strategy reduces the cognitive load of forming a judgement, but at the cost of losing some accuracy. Yet recent work has also shown that human heuristics are often the best that can be done in the presence of bounded computational resources, incomplete knowledge, ambiguous situations or limited decision time.3, 4 This has spurred research efforts to develop a computational scheme that incorporates what is known about the building blocks of the typical high-level functioning of the human brain.5 Organic Computing (OC) is a vision for future information processing systems intended to meet the challenges posed to designers and users by the current trajectory of developments in IT, such as ever increasing interconnectedness and computational power. The idea of OC is to tackle these upcoming demands by creating more lifelike (organic) computational systems endowed with abilities such as self-organization, self-configuration, self-repair and adaptation. Distributing computational intelligence through approaches such as self-organization relieves designers from having to exactly specify the low-level behaviour of a system in every possible situation. It also means the user can define just a few high-level goals, rather than having to manipulate many low-level parameters. Alongside this search for more efficient general computational algorithms and procedures, IT systems are also increasingly being required to interpret the behaviour of humans and users to operate effectively. One way to reduce the complexity of this task is by designing self-aware artificial systems capable of organizing their knowledge in a similar manner to humans. Such systems would therefore feature processes (e.g. learning and decision making) and cognitive building blocks (e.g. mental schemes and cognitive heuristics) similar to those that characterize human cognition. A system structured in this way would be inherently equipped with the ability to understand, mimic and interpret human behaviours and their underlying cognitive dynamics. With this goal in mind, as part of our work for the European Commission Seventh Framework Programme (FP7) project RECOGNITION, we have developed a three-part, or tripartite, model of human cognition for implementing self-awareness in content-centric technological systems.6 Our idea was to assemble a working computational framework inspired by human heuristics, based on the theoretical building blocks of high-level cognitive functions. The resultant model, illustrated in Figure 1, consists of a ‘perception’ module (A) that deals with external inputs and populates the internal context (thus corresponding to the unconscious knowledge humans gain by perceiving and attending to the world around them), a ‘reasoning’ module (B) that processes this information and performs actions accordingly, and a ‘learning’ module (C) that controls the other two modules and is responsible for their evolution and adaptation. Figure 1. The tripartite computational model is patterned after the three main elements of high-level human cognitive functions. 
Modules A and B (perception and reasoning) are composed of both schemes and heuristics. Schemes handle interactions with the outside world, such as receiving inputs or performing actions, while the heuristics are internal sets of rules used to control the schemes. Module C (learning) does not have any direct external interactions and so consists only of heuristics, which feed back into the other two modules. Schemes and heuristics are designed in a similar way. There is an activation pattern that needs to be matched with the context before a scheme or heuristic can be triggered. If the context later changes so that it no longer matches the activation pattern, that scheme or heuristic must cease. For example, when a module A (perception) scheme is activated, it typically collects information from the outside world, filters it and populates the context with this information. In the process of doing this, it may end up being deactivated while other schemes are activated in their turn. These may be other module A schemes that process further data from the outside world, or module B (reasoning) schemes that carry out actions. Since it is also possible for conflicting schemes to be activated at the same time, modules A and B both contain heuristics that help the system make a decision in such cases. The way this works is that, once a scheme has been activated, it is not automatically allowed to modify the context. Instead, there is a sort of competition among schemes, based on their level of ‘confidence’ (an internal evaluation of the effective ability to process the data) and past scores (measures of previous success in reaching goals). The relevance (A) and recognition (B) heuristics deal with this type of conflict resolution, and also promote ‘under-activated’ schemes (those that fall just beneath the activation threshold) if there is a time constraint for performing an action. The overall progress of the system is monitored by the goal-setting heuristic of module B, while the completion heuristic (of module B) has the task of recognizing when an action has ended (including the cases where an action is aborted due to insufficient time or resources). Finally, the evaluation heuristics in module C have the task of learning and optimizing the system, by collecting feedback, ‘replaying’ actions stored in memory (as in mental simulations) and imitating other systems. In summary, we have developed a tripartite model to computationally mimic high-level cognitive functions. It consists of three modules, each of which is composed of schemes and heuristics. The schemes deal with external data and actions, while heuristics govern the operation of the schemes. We believe that such a model can be deployed in a number of ways. For example, it could be used in designing a self-aware application that runs on a portable device, with the capability to self-assemble audio/video streams for purposes such as entertainment, rehabilitation of the elderly and distance learning. Such an application would receive suggestions (inferred from users' choices and their directly expressed preferences) from a central database, and exploit its knowledge of the context (time of day, location, user activity and so forth) to assemble an information stream in real time. Another possible application of this model is for addressing the problem of community detection (since, like the modern World Wide Web, human social networks have to be considered a continuum of nested communities whose boundaries are somewhat arbitrary) using decentralized algorithms and local information.7 We also envisage that our model could be used to study the role of awareness during a computer infection (e.g. contact avoidance, community choice and so forth).6 Our future work will focus on defining in more detail and expanding upon these practical applications of the tripartite model. References
Stay Informed
|
||||
|
||||