International Journal on Artificial Intelligence Tools

The essential parameters approach is a well-known technique in computer vision for recovering the motion and scene parameters from a sequence of images. This approach has long been considered as suboptimal because of the underestimation of the effect of noise and outliers. This paper re-evaluates this method because it is correctly classified as a structured Total Least Squares problem and proposes very robust linear neurons for its solution. Then, a novel neural network, CASEDEL EXIN, which exploits these neurons together with the case deletion diagnostics, is introduced. It is not only very robust (outlier rejection), but is also able to identify the outliers. This fact can be exploited, for example, to refine the image segmentation.

Research reported in this article is motivated, in part, by current U.S. military programs aimed at the development of efficient data integration and sensor management methods capable of handling large sensor suites and achieving robust target recognition performance in real time scenarios. Modern sensor systems have shown good recognition abilities against a few isolated targets. However, these capabilities decline steeply when multiple sensors are acting against large target groups under realistic conditions requiring dynamic allocation of the sensor resources and efficient on-line integration and disambiguation of multiple sensor outputs. Neural networks and other sensor integration technologies have been inspired by cognitive models attributing human perceptual integration to parallel processing and convergence of simultaneous data streams. This article explores a different model emphasizing serial processing and association of consecutive memory traces in the Long Term Memory (LTM) into a globally connected memory structure called a Virtual Associative Network (VAN). Information integration in VAN is called blending. Target representation is constructed dynamically from the segments of virtual net matched serially against the input segments in the Short Term Memory (STM). This article will elaborate the concept of blending, reference its biological foundations, explain the difference between information blending and conventional sensor fusion techniques, and demonstrate blending applications in a large scale sensor management task.

Artificial Intelligence focuses on the question of how to design system to exhibit intelligent behaviour in complex environments. Complex global behaviours can emerge from simple systems acting in a complex environment; however, this emergence requires that the systems' internal structure reflect essential structures in the environment. This paper examines the algebraic structure of a system's actions. We find that these actions often possess a self-similar local neighborhood structure that permits analysis and synthesis to be performed locally yet produce global, intelligent behaviours. A procedure for finding this local structure is presented, and illustrated with examples.

Intimate to the functioning and behavior of intelligent systems is the manner in which information is represented internally. The conventional approach to intelligent system design assumes a particular bias in the manner by which this information is represented. Typically, this is characterized by an "abstract" or "objective" design methodology which holds that intelligence is not a function of the physical nature of the system. Such an approach suffers from several shortcomings, most notably problems relating to scaling and complexity. Recent physiological research, however, has demonstrated that physical bodily form is a fundamental building block in the organization of mammalian cortical structures. Consequently, this article explores such a biologically motivated "subjective" or "egocentric" approach to system design, and demonstrates its utility in a simple robot arm control problem.

The robustness of bipedal walking robots can be enhanced by the use of adaptive control techniques. In this paper, we extend a previous control approach. "Virtual Model Control" (VMC) [6] to create "Adaptive Virtual Model Control" (AVMC). The adaptation compensates for external disturbances and unmodelled dynamics, enhancing robustness in the control of height, pitch, and forward speed. The state machine used to modulate the virtual model components and to select the appropriate virtual to physical transformations (as in traditional VMC) is also used to inform the adaptation about the robot's changing configuration.

The design procedure for AVMC is described in this paper and simulation results are presented for a planar walking biped.

Controlling the running movement of a machine with jointed legs is a difficult problem in robotics. The control of a single running stride can be broken down into three phases; takeoff, ballistic, and landing. This paper describes the development of an intelligent controller for the landing phase of a simulated jointed leg's running stride. The landing controller takes control when the tool touches the ground after the airborne ballistic phase of the stride. It lears from experience to control the leg so as to recover from the impact with the ground and to reposition the leg for the takeoff of the next running stride. Very accurate control is achieved even during the first attempt at the running stride.