Brian G. Woolley

On the deleterious effects of a priori objectives on evolution and representation

Evolutionary algorithms are often evaluated by measuring and comparing their ability to consistently reach objectives chosen a priori by researchers. Yet recent results from experiments without explicit a priori objectives, such as in Picbreeder and with the novelty search algorithm, raise the question of whether the very act of setting an objective is exacting a subtle price. Nature provides another hint that the reigning objective-based paradigm may be obfuscating evolutionary computations true potential; after all, many of the greatest discoveries of natural evolution, such as flight and human-level intelligence, were not set as a priori objectives at the beginning of the search. The dangerous question is whether such triumphs only result because they were not objectives. To examine this question, this paper takes the unusual experimental approach of attempting to re-evolve images that were already once evolved on Picbreeder. In effect, images that were originally discovered serendipitously become a priori objectives for a new experiment with the same algorithm. Therefore, the resulting failure to reproduce the very same results cannot be blamed on the evolutionary algorithm, setting the stage for a contemplation of the price we pay for evaluating our algorithms only for their ability to achieve preconceived objectives.

Evolving a single scalable controller for an octopus arm with a variable number of segments

While traditional approaches to machine learning are sensitive to highdimensional state and action spaces, this paper demonstrates how an indirectly encoded neurocontroller for a simulated octopus arm leverages regularities and domain geometry to capture underlying motion principles and sidestep the superficial trap of dimensionality. In particular, controllers are evolved for arms with 8, 10, 12, 14, and 16 segments in equivalent time. Furthermore, when transferred without further training, solutions evolved on smaller arms retain the fundamental motion model because they simply extend the general kinematic concepts discovered at the original size. Thus this work demonstrates that dimensionality can be a false measure of domain complexity and that indirect encoding makes it possible to shift the focus to the underlying conceptual challenge.

A novel human-computer collaboration: combining novelty search with interactive evolution

Recent work on novelty and behavioral diversity in evolutionary computation has highlighted the potential disadvantage of driving search purely through objective means. This paper suggests that leveraging human insight during search can complement such novelty-driven approaches. In particular, a new approach called novelty-assisted interactive evolutionary computation (NA-IEC) combines human intuition with novelty search to facilitate the serendipitous discovery of agent behaviors in a deceptive maze. In this approach, the human user directs evolution by selecting what is interesting from the on-screen population of behaviors. However, unlike in typical IEC, the user can now request that the next generation be filled with novel descendants. The experimental results demonstrate that combining human insight with novelty search not only finds solutions significantly faster and at lower genomic complexities than fully-automated processes guided purely by fitness or novelty, but it also finds solutions faster than the traditional IEC approach. Such results add to the evidence that combining human users and automated processes creates a synergistic effect in the search for solutions.

Real-time behavior-based robot control

Behavior-based systems form the basis of autonomous control for many robots, but there is a need to ensure these systems respond in a timely manner. Unexpected latency can adversely affect the quality of an autonomous system’s operations, which in turn can affect lives and property in the real-world. A robots ability to detect and handle external events is paramount to providing safe and dependable operation. This paper presents a concurrent version of a behavior-based system called the Real-Time Unified Behavior Framework, which establishes a responsive basis of behavior-based control that does not bind the system developer to any single behavior hierarchy. The concurrent design of the framework is based on modern software engineering principles and only specifies a functional interface for components, leaving the implementation details to the developers. In addition, the individual behaviors are executed by a real-time scheduler, guaranteeing the responsiveness of routines that are critical to the autonomous system’s safe operation. Experimental results demonstrate the ability of this approach to provide predictable temporal operation, independent of fluctuations in high-level computational loads.

Unified Behavior Framework for Reactive Robot Control

Behavior-based systems form the basis of autonomous control for many robots. In this article, we demonstrate that a single software framework can be used to represent many existing behavior based approaches. The unified behavior framework presented, incorporates the critical ideas and concepts of the existing reactive controllers. Additionally, the modular design of the behavior framework: (1) simplifies development and testing; (2) promotes the reuse of code; (3) supports designs that scale easily into large hierarchies while restricting code complexity; and (4) allows the behavior based system developer the freedom to use the behavior system they feel will function the best. When a hybrid or three layer control architecture includes the unified behavior framework, a common interface is shared by all behaviors, leaving the higher order planning and sequencing elements free to interchange behaviors during execution to achieve high level goals and plans. The framework’s ability to compose structures from independent elements encourages experimentation and reuse while isolating the scope of troubleshooting to the behavior composition. The ability to use elemental components to build and evaluate behavior structures is demonstrated using the Robocode simulation environment. Additionally, the ability of a reactive controller to change its active behavior during execution is shown in a goal seeking robot implementation.

Genetic evolution of hierarchical behavior structures

The development of coherent and dynamic behaviors for mobile robots is an exceedingly complex endeavor ruled by task objectives, environmental dynamics and the interactions within the behavior structure. This paper discusses the use of genetic programming techniques and the unified behavior framework to develop effective control hierarchies using interchangeable behaviors and arbitration components. Given the number of possible variations provided by the framework, evolutionary programming is used to evolve the overall behavior design. Competitive evolution of the behavior population incrementally develops feasible solutions for the domain through competitive ranking. By developing and implementing many simple behaviors independently and then evolving a complex behavior structure suited to the domain, this approach allows for the reuse of elemental behaviors and eases the complexity of development for a given domain. Additionally, this approach has the ability to locate a behavior structure which a developer may not have previously considered, and whose ability exceeds expectations. The evolution of the behavior structure is demonstrated using agents in the Robocode environment, with the evolved structures performing up to 122 percent better than one crafted by an expert.

Toward aircraft recognition with convolutional neural networks

We summarize the history and state of the art in Convolutional Neural Networks (CNNs), which constitute a significant advancement in pattern recognition. As a demonstration of capability, we address the problem of automatic aircraft identification during refueling approach. In this paper we describe the history of CNN development and provide a high level overview of the state of the art and a summary of leading CNN libraries with CUDA support. Finally, we demonstrate an application of CNN technology to autonomous aerial refueling and identify areas of follow-on research.

Mitigating the effects of boom occlusion on automated aerial refueling through shadow volumes

In-flight refueling of unmanned aerial vehicles (UAVs) is critical to the United States Air Force (USAF). However, the large communication latency between a ground-based operator and his/her remote UAV makes docking with a refueling tanker unsafe. This latency may be mitigated by leveraging a tanker-centric stereo vision system. The vision system observes and computes an approaching receiver’s relative position and orientation offering a low-latency, high frequency docking solution. Unfortunately, the boom – an articulated refueling arm responsible for physically pumping fuel into the receiver – occludes large portions of the receiver especially as the receiver approaches and docks with the tanker. The vision system must be able to compensate for the boom’s occlusion of the receiver aircraft. We present a novel algorithm for mitigating the negative effects of boom occlusion in stereo-based aerial environments. Our algorithm dynamically compensates for occluded receiver geometry by transforming the occluded areas into shadow volumes. These shadow volumes are then used to cull hidden geometry that is traditionally consumed, in error, by the vision processing and point registration pipeline. Our algorithm improves computer-vision pose estimates by 44% over a naïve approach without shadow volume culling.

Automated aerial refueling: Parallelized 3D iterative closest point: Subject area: Guidance and control

Vision-based automated aerial refueling requires a relative positioning solution capable of real time updates between the tanker and receiver. The processing time required for modern approaches are dominated by an iterative point alignment phase. This work presents an accelerated alignment variant which utilizes parallelization and Delaunay Triangulation to achieve real time estimates.

Parallelized Iterative Closest Point for Autonomous Aerial Refueling

The Iterative Closest Point algorithm is a widely used approach to aligning the geometry between two 3 dimensional objects. The capability of aligning two geometries in real time on low-cost hardware will enable the creation of new applications in Computer Vision and Graphics. The execution time of many modern approaches are dominated by either the k nearest neighbor search (kNN) or the point alignment phase. This work presents an accelerated alignment variant which utilizes parallelization on a Graphics Processing Unit (GPU) of multiple kNN approaches augmented with a novel Delaunay Traversal to achieve real time estimates.