Emad W. Saad

Neural network explanation using inversion

An important drawback of many artificial neural networks (ANN) is their lack of explanation capability [Andrews, R., Diederich, J., & Tickle, A. B. (1996). A survey and critique of techniques for extracting rules from trained artificial neural networks. Knowledge-Based Systems, 8, 373-389]. This paper starts with a survey of algorithms which attempt to explain the ANN output. We then present HYPINV, a new explanation algorithm which relies on network inversion; i.e. calculating the ANN input which produces a desired output. HYPINV is a pedagogical algorithm, that extracts rules, in the form of hyperplanes. It is able to generate rules with arbitrarily desired fidelity, maintaining a fidelity-complexity tradeoff. To our knowledge, HYPINV is the only pedagogical rule extraction method, which extracts hyperplane rules from continuous or binary attribute neural networks. Different network inversion techniques, involving gradient descent as well as an evolutionary algorithm, are presented. An information theoretic treatment of rule extraction is presented. HYPINV is applied to example synthetic problems, to a real aerospace problem, and compared with similar algorithms using benchmark problems.

Vehicle Swarm Rapid Prototyping Testbed

Increased levels of vehicle collaboration and auton omy are seen as a means to reduce overall mission completion costs while expanding mi ssion capabilities and increasing mission assurance for complex coupled system of systems. Systems health management technologies have made rapid advances that enable systems to know their own condition and capabilities, thus creating the opportunity for unprecedented lev els of adaptive control, real-time reconfiguration, and mission contingency management. Multi-agent task allocation and mission managements systems must account for vehicle- and system-level health-related issues to ensure that these systems are reliable an d cost effective to operate. Boeing’s Vehicle Swarm Technology Lab (VSTL), established in 2004, includes a 100’x50’x20’ testbed equipped with a vision-based motion capture indoor localization system. The testbed provides a cost-effective rapid prototyping capabil ity for integrating health-based adaptive control of subsystems, vehicle, mission, and swarms to guarantee top-level system-of-systems performance metrics. The lab’s heterogeneous fleet includes over 20 heterogeneous air vehicles, including VTOL and fixed wing, along with their ground stations and communication links in addition to heterogeneous ground vehicles and wall climbing robots. This paper discusses the Boeing VSTL design and capabilities, including the indoor localization system, multi-vehicle command and control (C2) and operator interface, realtime virtual environment, and health-based adaptive behaviors. The lab supports rapid prototyping and exploration of various multi-vehicl e operational concept of operations and missions including persistent surveillance, area se arch and tracking, and high density air traffic management. Additionally, the lab supports experimentation tasks for many other platform configuration and collaborative air, groun d, space, and maritime autonomous system of systems concepts.

Advanced neural network training methods for low false alarm stock trend prediction

Two possible neural network architectures for stock market forecasting are the time-delay neural network and the recurrent neural network. In this paper we explore two effective techniques for the training of the above networks: the conjugate gradient algorithm and multi-stream extended Kalman filter. We are particularly interested in limiting false alarms, which correspond to actual investment losses. Encouraging results have been obtained when using the above techniques.

Adaptive task allocation for search area coverage

Many operations require an area search area function, including search-and-rescue, surveillance, hazard detection, structures or sites inspection and agricultural spraying. Furthermore, these area search applications often involve varying vehicle and environmental conditions. This paper explores the problem of optimizing the behavior of a swarm of heterogeneous robotic vehicles executing a search area coverage task. Each vehicle is equipped with a sensing apparatus and the swarm must collectively explore an occluded environment to achieve a required probability of detection for each location in the search area. The problem is further complicated with the introduction of dynamic vehicle and environmental properties making adaptability a necessary requirement in order to achieve a high level of mission assurance using unmanned vehicles. Novel methods for search space decomposition and task allocation are presented, with simulated and real-world results utilizing the Boeing Vehicle Swarm Technology Laboratory.

Memetic Mission Management

This paper presents novel area coverage algorithms that have been validated using Boeing VSTL hardware. Even though the multi-vehicle search area coverage problem is large and complex, several new memetic computing methods have been presented that decompose, allocate and optimize the exploration of a search area for multiple heterogeneous vehicles. These new methods were shown to have good performance and quality, and as they are defined in a general way, these methods are applicable to many other problem domains. The methods have been combined into a mission-planner architecture that is able to adaptively control the behavior of multiple vehicles with dynamic vehicle capabilities and environments for mission assurance. The topic of mission-planning architectures and optimization of swarms of autonomous vehicles is a young and exciting field with many opportunities for research. More computationally efficient methods for decomposition may be useful, as well as the application of next-generation meta-learning architectures for path planning. In addition to the existing collision avoidance, path de-confliction during planning can improve safety and efficiency.

Memetic Mission Management [Application Notes]

This paper presents novel area coverage algorithms that have been validated using Boeing VSTL hardware. Even though the multi-vehicle search area coverage problem is large and complex, several new memetic computing methods have been presented that decompose, allocate and optimize the exploration of a search area for multiple heterogeneous vehicles. These new methods were shown to have good performance and quality, and as they are defined in a general way, these methods are applicable to many other problem domains. The methods have been combined into a mission-planner architecture that is able to adaptively control the behavior of multiple vehicles with dynamic vehicle capabilities and environments for mission assurance. The topic of mission-planning architectures and optimization of swarms of autonomous vehicles is a young and exciting field with many opportunities for research. More computationally efficient methods for decomposition may be useful, as well as the application of next-generation meta-learning architectures for path planning. In addition to the existing collision avoidance, path de-confliction during planning can improve safety and efficiency.

Query-based learning for aerospace applications

Models of real-world applications often include a large number of parameters with a wide dynamic range, which contributes to the difficulties of neural network training. Creating the training data set for such applications becomes costly, if not impossible. In order to overcome the challenge, one can employ an active learning technique known as query-based learning (QBL) to add performance-critical data to the training set during the learning phase, thereby efficiently improving the overall learning/generalization. The performance-critical data can be obtained using an inverse mapping called network inversion (discrete network inversion and continuous network inversion) followed by oracle query. This paper investigates the use of both inversion techniques for QBL learning, and introduces an original heuristic to select the inversion target values for continuous network inversion method. Efficiency and generalization was further enhanced by employing node decoupled extended Kalman filter (NDEKF) training and a causality index (CI) as a means to reduce the input search dimensionality. The benefits of the overall QBL approach are experimentally demonstrated in two aerospace applications: a classification problem with large input space and a control distribution problem.

Aggressive navigation using high-speed natural feature point tracking

Presently, most autonomous aerial vehicles rely on satellite navigation such as GPS to sense their position in the earth reference frame. However, reliance on GPS restricts the vehicle to missions where GPS signals are readily received. Motion capture systems are capable of indoor localization but require large infrastructure and are prone to occlusion. To overcome these restrictions, a self-contained high-speed vision system was developed at the University of Tokyo in collaboration with Boeing Research & Technology. The system has been flight tested and shown to be capable of drift-free position and attitude estimates without any reliance on GPS signals. Furthermore, the positional accuracy and update rate is at least one order of magnitude superior to that of uncorrected GPS. The vision system combines a high-speed camera with a lightweight computer and power supply into a self-contained computer-vision package. The computer processes the incoming image stream with a modified version of the University of Oxford Parallel Tracking and Mapping (PTAM) SLAM algorithm. Using this algorithm, the location and pose of the camera (and the MA V it is attached to) is estimated as it moves through space by mapping natural features as they first appear and by tracking those features as they move through or reappear in the cameras view. Our vision system was demonstrated using a hexacopter test bed. In a pair of experiments, the hexacopter was able to autonomously repeat a circuit of takeoffs and landings at predetermined separated sites using only MEMs gyro sensors and our vision system. One experiment was performed inside Boeings motioncapture equipped Collaborative Systems Laboratory (CSL) to prove independence from GPS and to measure the accuracy of the vision system. The hexacopter performed 5 circuits of the navigation task over an area approximately 8 x 8 m at an altitude of approximately 2 m. The vision systems camera was set to provide an image stream of 640 x 480 pixel resolution at 50 Hz. Upon comparison with motion capture data, position estimates from the vision system were shown to be free of drift, with an average error of 2.2 cm and a maximum error of 9.7 cm when the vision system coordinate frame was optimally aligned to the motion capture coordinate frame. A second experiment was performed in an open outdoor area, allowing for safe execution of more aggressive maneuvering. In this experiment, the vision systems camera was set to provide an image stream of 320 x 240 pixel resolution at 120 Hz. This experiment demonstrated the ability to perform takeoffs, landings and transit at higher speeds than was demonstrated in the indoor experiment.

Predictive head tracking for virtual reality

In virtual reality (VR), head movement is tracked through inertial and optical sensors. Computation and communication times result in delays between measurements and updating of the new frame in the head mounted display(HMD). These delays result in problems, including motion sickness. We use recurrent and time delay neural networks to predict the head location and use it to calculate the new frame. A predictability analysis is used in designing the prediction system.

Sensitivity study for feature-based monocular 3D SLAM

Advances in cameras and computing hardware, both in terms of performance and miniaturization, has made vision-based localization a feasible sensor for aerial vehicles. In GPS deprived environments or scenarios where the resolution of GPS is not sufficient, such a sensor presents an attractive alternative. Vision-based position sensors typically estimate their pose by tracking natural features in the environment, while at the same time creating a map of those features. This process is referred to as simultaneous localization and mapping (SLAM), and it employs several sub-processes, such as feature detection and description, map generation, feature mapping, and optimization, each of which is subject to a large number of parameters. Due to the complexity of the problem, finding a satisfactory parameter setting can be a tedious task. In this paper we investigate the effects of each parameter in the context of SLAM. As an example we use the PTAM (Parallel Tracking and Mapping) algorithm from the University of Oxford. The results of this sensitivity study indicate which parameters are most influential in achieving good tracking performance and also show suitable ranges for each parameter. This information can be used to expedite discovery of a satisfactory parameter setting for a new environment.