On-line Adaptation Capabilities Research Articles

A Pilot Study of Simulation-Free Hippocampal-Avoidance Whole Brain Radiation Therapy Using Diagnostic MRI-Based and Online Adaptive Planning

We aimed to demonstrate the clinical feasibility and safety of simulation-free hippocampal avoidance whole brain radiation therapy (HA-WBRT) in a pilot study (National Clinical Trial 05096286). Ten HA-WBRT candidates were enrolled for treatment on a commercially available computed tomography (CT)-guided linear accelerator with online adaptive capabilities. Planning structures were contoured on patient-specific diagnostic magnetic resonance imaging (MRI), which were registered to a CT of similar head shape, obtained from an atlas-based database (AB-CT). These patient-specific diagnostic MRI and AB-CT data sets were used for preplan calculation, using NRG-CC001 constraints. At first fraction, AB-CTs were used as primary data sets and deformed to patient-specific cone beam CTs (CBCT) to give patient-matched density information. Brain, ventricle, and brain stem contours were matched through rigid translation and rotation to the corresponding anatomy on CBCT. Lens, optic nerve, and brain contours were manually edited based on CBCT visualization. Preplans were then reoptimized through online adaptation to create final, simulation-free plans, which were used if they met all objectives. Workflow tasks were timed. In addition, patients underwent CT-simulation to create immobilization devices and for prospective dosimetric comparison of simulation-free and simulation-based plans. Median time from MRI importation to completion of "preplan" was 1 weekday (range, 1-4). Median on-table workflow duration was 41 minutes (range, 34-70). NRG-CC001 constraints were achieved by 90% of the simulation-free plans. One patient's simulation-free plan failed a planning target volume coverage objective (89% instead of 90% coverage); this was deemed acceptable for first-fraction delivery, with an offline replan used for subsequent fractions. Both simulation-free and simulation CT-based plans otherwise met constraints, without clinically meaningful differences. Simulation-free HA-WBRT using online adaptive radiation therapy is feasible, safe, and results in dosimetrically comparable treatment plans to simulation CT-based workflows while providing convenience and time savings for patients.

A Pilot Study of Simulation-Free Hippocampal-Avoidance Whole Brain Radiotherapy Using Diagnostic MR-Based and Online Adaptive Planning

We aimed to demonstrate the clinical feasibility and safety of a simulation-free hippocampal avoidance whole brain radiotherapy (HA-WBRT) workflow in a Phase I clinical trial (NCT05096286). Feasibility was defined as successful completion of the simulation-free HA-WBRT workflow through treatment delivery in at last 70% of treated patients. Ten candidates for HA-WBRT were enrolled for treatment on a ring gantry CT-guided Linac with online adaptive capabilities. Structures were contoured on the diagnostic brain MRI, which was then registered to a separate head computed tomography (CT) of similar head shape, obtained from an atlas-based database. A HA-WBRT "pre-plan" was generated using the atlas-based CT (AB-CT) and the NRG-CC001 constraints. At first fraction, the AB-CT was used as the primary dataset and deformed to the patient's cone-beam CT (CBCT) for dose calculation. The brain, ventricles, and brainstem contours were matched through rigid translation and rotation to the corresponding anatomy on the CBCT to aid in alignment, given the differences in rotational head positioning from diagnostic MRI to CBCT setup. Lastly, the lens, optic nerves, and brain contours were manually edited based on CBCT visualization. Plans were then optimized, and the adaptive plan was chosen for treatment if the plan met all objectives. Workflow tasks were timed. In addition, conventional plans using patients' sim CTs were created for each patient for the purpose of prospective dosimetric comparison. The dosimetric parameters were compared for each patient between the delivered sim-free plan and the conventional sim CT plan using the sign test via statistical software, with p<.05 indicating significance. Median time from approved sim order to first fraction was 4 days (range: 2-7); median time in room (door-to-door) was 49 minutes (range: 35-70). All patients successfully completed all ten fractions and 90% of the simulation-free radiation plans met all NRG-CC001 constraints. For one patient, the sim-free plan at fraction one failed the planning target volume (PTV) coverage objective (coverage of 89%); this was deemed acceptable for delivery by the treating radiation oncologist. An offline replan was then performed to meet NRG-CC001 constraints and used for the subsequent nine fractions. There was no clinically meaningful difference in dosimetric constraints between the sim-free plan (calculated on AB-CT) and conventional CT sim plan. Statistically, the sim-free plans provided improved PTV coverage to higher doses compared to the conventional plans (Table). At a median follow-up of 43 days (range: 9 -280), the intracranial progression-free survival rate was 90%. Simulation-free HA-WBRT is feasible, results in plans that are dosimetrically comparable to conventional CT sim workflows and succeeds in decreasing time to initiation of HA-WBRT by at least 50%. Further studies with a larger cohort are warranted to optimize the workflow.

Safe Active Dynamics Learning and Control: A Sequential Exploration–Exploitation Framework

Safe deployment of autonomous robots in diverse scenarios requires agents that are capable of efficiently adapting to new environments while satisfying constraints. In this article, we propose a practical and theoretically justified approach to maintain safety in the presence of dynamics uncertainty. Our approach leverages Bayesian meta-learning with last-layer adaptation. The expressiveness of neural-network features trained offline, paired with efficient last-layer online adaptation, enables the derivation of tight confidence sets, which contract around the true dynamics as the model adapts online. We exploit these confidence sets to plan trajectories that guarantee the safety of the system. Our approach handles problems with high dynamics uncertainty, where reaching the goal safely is potentially initially infeasible, by first <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">exploring</i> to gather data and reduce uncertainty, before autonomously <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">exploiting</i> the acquired information to safely perform the task. Under reasonable assumptions, we prove that our framework guarantees the high-probability satisfaction of all constraints at all times jointly, i.e., over the total task duration. This theoretical analysis also motivates two regularizers of last-layer meta-learning models that improve online adaptation capabilities as well as performance by reducing the size of the confidence sets. We extensively demonstrate our approach in simulation and on hardware.

Variable Impedance Control and Learning-A Review.

Robots that physically interact with their surroundings, in order to accomplish some tasks or assist humans in their activities, require to exploit contact forces in a safe and proficient manner. Impedance control is considered as a prominent approach in robotics to avoid large impact forces while operating in unstructured environments. In such environments, the conditions under which the interaction occurs may significantly vary during the task execution. This demands robots to be endowed with online adaptation capabilities to cope with sudden and unexpected changes in the environment. In this context, variable impedance control arises as a powerful tool to modulate the robot's behavior in response to variations in its surroundings. In this survey, we present the state-of-the-art of approaches devoted to variable impedance control from control and learning perspectives (separately and jointly). Moreover, we propose a new taxonomy for mechanical impedance based on variability, learning, and control. The objective of this survey is to put together the concepts and efforts that have been done so far in this field, and to describe advantages and disadvantages of each approach. The survey concludes with open issues in the field and an envisioned framework that may potentially solve them.

A New LMS Algorithm Based Deadtime Compensation Method for PMSM FOC Drives

This paper presents a new least-mean-square (LMS) algorithm based deadtime compensation method to suppress the current distortion in permanent-magnet synchronous motor (PMSM) field-oriented control (FOC) drives. Compared to conventional average value compensations, the proposed method is robust to switching device parameter variations thanks to the online adaptation capability of the LMS algorithm. Similarly, the disturbance observer compensators are also immune to switching device parameter variations; however, varying motor parameters degrade their compensation performance. Without prior knowledge of switching device or motor parameters, the proposed method can directly reduce the deadtime current harmonics by generating compensation voltage references. In addition, the proposed method is easy to implement since it does not require voltage errors estimation or current harmonics extraction, which are necessary for disturbance observer and adaptive filter based methods. The proposed method is tested on a 2.5-kW voltage-source inverter PMSM drive controlled by an FOC algorithm. Its effectiveness is validated by both experimental results and spectrum analysis.

Design of a Gait Phase Recognition System That Can Cope With EMG Electrode Location Variation

Electromyogram (EMG) signal-based gait phase recognition for walking-assist devices warrants much attention in human-centered system design as it well exemplifies human-in-the-loop control where the system's prediction directly affects subsequent walking motion. Since walking motion poses considerable variations in electrode placement, performance reliability of such systems is contingent on a combination of electrode montage and a feature extraction method that takes into account underlying physiological factors of peripheral muscles where electrodes are placed. In many practical applications, however, proper consideration of effects of the electrode location variation on performance reliability of the system has received scant empirical attention. Here, based on a user-centered design principle, we establish a gait phase recognition system that is capable of rigidly controlling ill effects due to this covariate by carrying out a large-scale analysis that combines statistical, model-based, and empirical approaches. In doing so, we have developed a special sensing suit for the control of electrode placement and a reliable data acquisition. We then have conducted a nonparametric statistical analysis on class separability values of thirty types of EMG feature sets, followed by a model-based analysis to address the tradeoff between class separability and dimensionality. To further address the issue of how these results generalize to independent systems and data sets, we have carried out an empirical performance assessment over six classification methods. First, the two feature types, Integral of Absolute Value and Histogram, and a combination of the two are shown to be robust against electrode location variations while providing a firm performance guarantee. Second, system organization scenarios are presented on a case-by-case basis, allowing us to trade off system complexity for on-line adaptation capability. Collectively, our integrated analysis lends itself to formulating a guideline for design of highly reliable EMG signal-based walking assistant systems in a variety of smart home scenarios.

Operator-Splitting and Gradient Methods for Real-Time Predictive Flight Control Design

Despite their ability to operate on the limits of performance, handle multivariable and nonlinear systems, and offer online adaptation and reconfiguration capabilities, model predictive control approaches to aerospace applications suffer from limitations related to the online computational burden and complexity of the underlying optimization problem. This paper focuses on quadratic programming (QP) formulations that represent certain types of predictive flight control problems, and proposes a parallelizable QP solver based on operator-splitting and fast gradient methods. The presented methodology and solution approach promise real-time implementation of QP-based predictive flight control schemes on future embedded platforms. This paper also provides formal analysis and guidelines on how to reshape the feasible region of the model predictive control problem at each problem instance to ensure recursive feasibility and closed-loop stability. Finally, simulation results for the longitudinal control of an Airbus passenger aircraft are presented to show how the obtained computable certificates can be simplified in practice and bring the proposed approach a step closer to future on-board real-time implementation.

On the Integration of Adaptive and Interactive Robotic Smart Spaces

AbstractEnabling robots to seamlessly operate as part of smart spaces is an important and extended challenge for robotics R&amp;D and a key enabler for a range of advanced robotic applications, such as AmbientAssisted Living (AAL) and home automation. The integration of these technologies is currently being pursued from two largely distinct view-points: On the one hand, people-centred initiatives focus on improving the user’s acceptance by tackling human-robot interaction (HRI) issues, often adopting a social robotic approach, and by giving to the designer and - in a limited degree – to the final user(s), control on personalization and product customisation features. On the other hand, technologically-driven initiatives are building impersonal but intelligent systems that are able to pro-actively and autonomously adapt their operations to fit changing requirements and evolving users’ needs, but which largely ignore and do not leverage human-robot interaction and may thus lead to poor user experience and user acceptance. In order to inform the development of a new generation of smart robotic spaces, this paper analyses and compares different research strands with a view to proposing possible integrated solutions with both advanced HRI and online adaptation capabilities.

Fault-tolerant adaptive control of nonlinear base-isolated buildings using EMRAN

This paper presents a direct adaptive fault-tolerant neural control scheme for the active control of nonlinear hysteretic base-isolated buildings using the recently developed Extended Minimal Resource Allocation Network (EMRAN). EMRAN is a learning algorithm in which the structure of the neural controller is adapted on-line based on the input–output data. EMRAN starts with no hidden neurons and calculates the number of hidden neurons based on growing/pruning criteria. If the criteria are not met, then the parameters of the network are adjusted using an Extended Kalman Filter (EKF). The constants associated with the growing/pruning criteria and EKF are estimated using Genetic Algorithm (GA) optimization. The advantage of the proposed control architecture is its ability to learn on-line with no a priori training. Most of the existing studies in structural control using neural networks require computationally intensive off-line training. Consequently, once the network parameters are learnt, the parameters remain fixed. Such procedures require an accurate mathematical model of the system. These issues are addressed in the current controller scheme by utilizing the on-line adaptation capabilities of the neural networks. The advantages of on-line adaptation are demonstrated using the controller’s capability to handle actuator failures and system uncertainties. Performance of the proposed control scheme is evaluated using the recently developed nonlinear three-dimensional base-isolated benchmark structure incorporating lateral–torsional superstructure behavior and the biaxial interaction of the nonlinear bearings in the isolation layer. Results show that the proposed controller scheme can achieve the desired performance objectives under both partial actuator failure conditions and large uncertainties associated with the system’s parameters.

가버 웨이블릿 신경망 기반 적응 표정인식 시스템

본 논문에서는 6개의 특징점을 이용하는 가버 웨이블릿 신경망 기반 적응 표정인식 시스템을 제안한다. 특징 추출부를 포함하는 초기 네트워크의 구성은 Levenberg-Marquardt 기반의 학습방법이 사용되며, 따라서 특징 추출부 결정에 있어서 경험적 요소를 배재시킬 수 있다. 또한 새로운 사용자에 대한 적응 네트워크를 구성하기 위해서 개선된 보상함수를 가지는 Q-학습과, 비지도 퍼지 신경망 모델을 사용하였다. Q-학습을 통해서는 개인 사용자에 대해 분리도가 좋은 특징벡터를 얻을 수 있는 가버필터 세트를 얻을 수 있으며, 퍼지 신경망을 통해서는 사용자의 얼굴변화에 맞게 인식기를 변화시킬 수 있다. 따라서 제안된 시스템은 사용자의 얼굴변화를 따라갈 수 있는 좋은 적응 성능을 보이고 있다. In this paper, adaptive Facial Emotional Recognition system based on Gabor Wavelet Neural Network, considering six feature Points in face image to extract specific features of facial expression, is proposed. Levenberg-Marquardt-based training methodology is used to formulate initial network, including feature extraction stage. Therefore, heuristics in determining feature extraction process can be excluded. Moreover, to make an adaptive network for new user, Q-learning which has enhanced reward function and unsupervised fuzzy neural network model are used. Q-learning enables the system to ge optimal Gabor filters' sets which are capable of obtaining separable features, and Fuzzy Neural Network enables it to adapt to the user's change. Therefore, proposed system has a good on-line adaptation capability, meaning that it can trace the change of user's face continuously.

Blind Source Recovery in a State-Space Famework: Algorithms for Static and Dynamic Environments

This paper describes a generalized state space Blind Source Recovery (BSR) framework obtained by using the Kullback-Lieblar divergence as a performance functional and the application of optimization theory under the constraints of a feedforward state space structure. Update laws for both the non-linear and the linear dynamical systems have been derived for the domain of dynamic blind source recovery along both ordinary stochastic gradient and the Riemannian contra-variant gradient directions. The choice of the rich state space demixing network structure allows for the development of potent learning rules, capable of handling most filtering paradigms and convenient extension to non-linear models. Some particular filtering cases are subsequently derived from this general structure and are compared with material in the recent literature. Some of this reported work has also been implemented in dedicated hardware/software. An illustrative simulation example has been presented to demonstrate the online adaptation capabilities of the proposed algorithms.

An adaptive window width/center adjustment system with online training capabilities for MR images

Adaptive and automatic adjustment of the display window parameters for magnetic resonance images under different viewing conditions is a challenging problem in medical image perception. An adaptive hierarchical neural network-based system with online adaptation capabilities is presented to achieve this goal in this paper. The online adaptation capabilities are primarily attributed to the use of the hierarchical neural networks and the development of a new width/center mapping algorithm. The large training image set is hierarchically organized for efficient user interaction and effective re-mapping of the width/center settings. The width/center mapping functions are estimated from the new user-adjusted width/center values of some representative images by using a global spline function for the entire training images as well as a first-order polynomial function for each selected image sequence. The hierarchical neural networks are then re-trained for the new training data set after this mapping process. The proposed automatic display window parameter adjustment system is implemented as a program on a personal computer for testing its adaptation performance. Experimental results show that the proposed system can successfully adapt its parameter adjustment on a variety of MR images after user re-adjustment and re-training of neural networks. This demonstrates the effective adaptation capabilities of the proposed system based on the framework of training data mapping and neural network re-training.

Automatization of a penicillin production process with soft sensors and an adaptive controller based on neuro fuzzy systems

This paper addresses the automatization of a penicillin production process with the development of soft sensors as well as Internal Model Controllers (IMC) for a penicillin fermentation plant using modules based on FasArt and FasBack neuro-fuzzy systems. While soft sensors are intended to aid the human supervision of the process currently being conducted at pilot plants, the proposed controller will make the automatization feasible and eliminate the need for human operator. FasArt and FasBack feature fast stable learning and good MIMO identification, which makes them suitable for the development of adaptive controllers and soft sensors. In this paper, these modules are evaluated by training the neuro-fuzzy systems first on simulated data and then applying the resulting IMC controllers to a simulated plant. Moreover, training the systems on data coming from a real pilot plant, and evaluating the controller performance on the same real plant. Results show that the trend of reference is captured, thus allowing high penicillin production. Moreover, soft sensors derived for biomass, viscosity and penicillin are very accurate. In addition, on-line adaptive capabilities were implemented and tested with FasBack, since this system presents learning guided by error minimization as new data samples arrive. With these features, adaptive IMC controllers can be implemented and are helpful when dynamics have been poorly learned or the plant parameters vary with time, since the performance of static models and controllers can be improved through adaptation.

Application of Radial Basis Function Neural Networks to a Greenhouse Inside Air Temperature Model

The problem of the adequacy of radial basis function neural networks to model the inside air temperature as a-function of the outside air temperature and solar radiation, and the inside relative humidity in an hydroponic greenhouse is addressed. This type of network is structurally simple and suitable to be integrated in real-time greenhouse environmental control systems. Due to the time variability of the process, training methods with on-line adaptation capabilities are neededThree of such methods are analysed in terms of fitness and network size. The model-predictive outputs obtained showed very close fittings to the measured values.

ZPM Compensation and Impedance Control for Improving Walking Stability of Biped Robots

This paper proposes an adaptive trajectory generation strategy of using on-line ZMP information and an impedance control method for biped robots. Since robots experience various disturbances during their locomotion, their walking mechanism should have the robustness against those disturbances, which requires an on-line adaptation capability. In this context, an on-line trajectory planner is proposed to compensate the required moment for recovering stability. The ZMP equation and sensed ZMP information are used in this trajectory generation strategy. In order to control a biped robot to be able to walk stably, its controller should guarantee stable footing at the moment of feet contacts with the ground as well as maintaining good trajectory tracking performance. Otherwise, the stability of robot will be significantly compromised. To reduce the magnitude of an impact and guarantee a stable footing when a foot contacts with the ground, this paper. proposes to increase the damping of the leg drastically and to modify the reference trajectory of the leg. In the proposed control scheme, the constrained leg is controlled by impedance control using the impedance model with respect to the base link. Computer simulations performed with a 3-dof environment model that consists of combination of a nonlinear and linear compliant contact model show that the proposed controller performs well and that it has robustness against unknown uneven surface. Moreover, the biped robot with the proposed trajectory generator can walk even when it is pushed with a certain amount of external force.

10.1162/jmlr.2003.4.7-8.1411

Blind Source Recovery (BSR) denotes recovery of original sources/signals from environments that may include convolution, temporal variation, and even nonlinearity. It also infers the recovery of sources even in the absence of precise environment identifiability. This paper describes, in a comprehensive fashion, a generalized BSR formulation achieved by the application of stochastic optimization principles to the Kullback-Liebler divergence as a performance functional subject to the constraints of the general (i.e., nonlinear and time-varying) state space representation. This technique is used to derive update laws for nonlinear time-varying dynamical systems, which are subsequently specialized to time-invariant and linear systems. Further, the state space demixing network structures have been exploited to develop learning rules, capable of handling most filtering paradigms, which can be conveniently extended to nonlinear models. In the special cases, distinct linear state-space algorithms are presented for the minimum phase and non-minimum phase mixing environment models. Conventional (FIR/IIR) filtering models are subsequently derived from this general structure and are compared with material in the recent literature. Illustrative simulation examples are presented to demonstrate the online adaptation capabilities of the developed algorithms. Some of this reported work has also been implemented in dedicated hardware/software platforms.

On-Line Multivariable Intelligent Control of an AUV

The design of control systems for autonomous underwater vehicles is a challenging task owing to their nonlinear characteristics and the severe cross-coupling which exists between the principal axes. In this paper, a novel multivariate neurofuzzy autopilot that possesses on-line adaptive capabilities is described and used to control simultaneously the yaw and roll channels of an autonomous underwater vehicle model. Simulation results are presented which demonstrate the effectiveness of the autopilot design.

Control of the Penicillin Production with Adaptive IMC Using Fuzzy Neural Networks

This paper introduces the use adaptation in IMC strategy for the control of a simulated penicillin plant. The plant model and control modules are built using FasBack neuro-fuzzy system, featuring fast stable learning guided by matching and error minimisation and good identification performance. Control results show good general performance both in the nominal case and in the presence of noise. FasBack on-line adaptation capabilities are used to develop an adaptive IMC, which shows to improve performance in realistic cases of time varying parameters. Furthermore, real data coming from pilot plants are used to train fuzzy neural networks with satisfactory identification results, and obtained modules are used within IMC with similar results to those build from simulated data.

Intentional On-line Adaptation of Rhythmic Movements during a Hyper- to Microgravity Change

The intentional on-line adaptive capabilities of human movements during continuous variation in gravitational force were investigated. Subjects performed rhythmic forearm movements in the gravitational plane during parabolic flight maneuvers that induced a continuous change from 1.8 G to 0 G over a period of 2.3 s. During the initial plateau of hypergravity, subjects produced movements at two frequencies, with and without space constraints. Afterward, they were faced with the drop in gravity, during which they were instructed either to let the movement evolve freely while maintaining the initial frequency (time-constrained task) or to intentionally maintain the frequency, amplitude, and forearm center of oscillation (time/space-constrained task). The results showed (a) a reduced angle for the forearm center of oscillation and maintenance of movement amplitude in the time-constrained task, (b) a change from an in-phase to an out-of-phase biceps/triceps activation pattern regardless of the task, and (c) an earlier occurrence of this change in the time/space-constrained task, impeding the spontaneous forearm rise. These results are discussed in the perspective of the λ model.

Dynamic control of a robot arm using CMAC neural networks

Neural identification and control techniques are well-suited to the problem of controlling robot dynamics. This paper describes the use of CMAC networks for the adaptive dynamic control of an orange-harvesting robot. Among the various neural-network paradigms available, the CMAC model was chosen in this case because of its fast convergence and on-line adaptation capability. The solution of this dynamic control problem with CMAC is an encouraging demonstration of “experience-based”, as opposed to model-based, control techniques and is a good example of the use of on-line learning in adaptive neural control.