Soft Computing Approaches for Intelligence Augmentation of Human Centered Systems

Human Centered Systems are systems intended to interact directly with humans. System performance, user included, hinges on the system's ability to provide an efficient communication channel between the user and the underlying machine processes. Special attention must be given to understand how the behavior and changes in the parameters of such systems are perceived by humans. To achieve a high level of performance, Human Centered Systems need intelligent interfaces---carefully designed interfaces that respect human perceptual capabilities and that have the ability to adapt to preference variations among users. Intelligence can be embedded into a human centered system in the following ways: by design, by enabling adaptation through user parameters, by enabling adaptation through self-learning, or by some combination of these methods. Because of the inherent imprecision and uncertainty in dealing with human users, we propose using Soft Computing-based methodologies and components in the design and architecture of Intelligent Human Centered Systems.

Soft Computing [52] is a consortium of methodologies whose essence is to exploit the pervasive imprecision of the real world, tolerance for imprecision, uncertainty and partial truth to achieve tractability, robustness, low solution cost, and better rapport with reality. The principal complementary and synergistic members of this consortium are fuzzy logic(FL), neurocomputing (NC), genetic algorithms (GA), and probabilistic reasoning (PR), with the latter subsuming evidential reasoning, belief networks, management of uncertainty, and parts of machine learning theory.

Our present work has focused on the development of hybrid soft computing techniques. Among the hybrid combinations we are researching include neuro-fuzzy, fuzzy-genetic, and neuro-fuzzy-genetic. Neuro-fuzzy systems are employed for rule induction from observations[53]. This problem is extremely relevant to human behavior modeling and human utility modeling in the context of human centered system design. We also exploring the use of fuzzy and neuro techniques applied to genetic algorithms. In this scenario, genetic algorithms are hybridized to include other adaptive mechanisms to improve system and learning performance [54]. Among the directions we intend to pursue is to use genetic algorithms to design systems for multiattribute tasks. The techniques we plan to develop are aimed at providing the user/designer with a set of solutions that represent the set of best alternatives as opposed to a single solution. From this set, the user can better understand the trade-offs between the objectives and make a more informed choice. Other directions we intend to pursue are in the area of Dynamic Switching Fuzzy Systems[55]. In such systems, fuzzy rule-bases are tuned online by modulating the fuzzy operators used to perform aggregation.

Soft Computing for the Design and User Adaptation of Human Centered Systems
User adaptation involves allowing the user to change the system behavior through a set of parameters, which are independently connected to some perceptual axis. To effect a perceptual change along some perceptual axis requires coordinating the parameter changes. What is needed are techniques to build mappings from a perceptual space to the underlying algorithm parameter space. We propose using soft computing components to represent the mapping from perceptual space to the underlying system parameter space. For example, if the mapping can be described by a collection of fuzzy if-then rules and the user is able to change the definition of the linguistic variables used in the rules, a range of different, but related, behaviors can be achieved by the system.

Soft Computing for Self Adaptation in Human Centered Systems
Self-adaptation involves enabling the interface to adapt system parameters based on direct or indirect user feedback. We propose using reinforcement learning methodologies and fuzzy systems to construct models of human preferences and human utility functions. The use of fuzzy systems to represent the knowledge has several advantages: prior knowledge can be encoded into the system using simple if-then rules, prior knowledge can be refined and augmented using fuzzy system learning techniques, and, given the proper constraints, the potential for acquired knowledge to be further interrogated and manipulated by humans following learning remains high.

Showing architecture for intelligent augmentation

Top