[ { "id": "authors:2q00r-v6053", "collection": "authors", "collection_id": "2q00r-v6053", "cite_using_url": "https://resolver.caltech.edu/CaltechAUTHORS:20190410-120654801", "type": "conference_item", "title": "End-to-End Safe Reinforcement Learning through Barrier Functions for Safety-Critical Continuous Control Tasks", "author": [ { "family_name": "Cheng", "given_name": "Richard", "clpid": "Cheng-Richard-P" }, { "family_name": "Orosz", "given_name": "G\u00e1bor", "orcid": "0000-0002-9000-3736", "clpid": "Orosz-G" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "Burdick", "given_name": "Joel W.", "clpid": "Burdick-J-W" } ], "abstract": "Reinforcement Learning (RL) algorithms have found limited success beyond simulated applications, and one main reason is the absence of safety guarantees during the learning process. Real world systems would realistically fail or break before an optimal controller can be learned. To address this issue, we propose a controller architecture that combines (1) a model-free RL-based controller with (2) model-based controllers utilizing control barrier functions (CBFs) and (3) on-line learning of the unknown system dynamics, in order to ensure safety during learning. Our general framework leverages the success of RL algorithms to learn high-performance controllers, while the CBF-based controllers both guarantee safety and guide the learning process by constraining the set of explorable polices. We utilize Gaussian Processes (GPs) to model the system dynamics and its uncertainties. \n\nOur novel controller synthesis algorithm, RL-CBF, guarantees safety with high probability during the learning process, regardless of the RL algorithm used, and demonstrates greater policy exploration efficiency. We test our algorithm on (1) control of an inverted pendulum and (2) autonomous car-following with wireless vehicle-to-vehicle communication, and show that our algorithm attains much greater sample efficiency in learning than other state-of-the-art algorithms and maintains safety during the entire learning process.", "doi": "10.48550/arXiv.1903.08792", "publisher": "arXiv", "publication_date": "2019-02" }, { "id": "authors:3tsrf-23g13", "collection": "authors", "collection_id": "3tsrf-23g13", "cite_using_url": "https://resolver.caltech.edu/CaltechAUTHORS:20171128-230807299", "type": "conference_item", "title": "Measuring the Robustness of Neural Networks via Minimal Adversarial Examples", "author": [ { "family_name": "Dathathri", "given_name": "Sumanth", "clpid": "Dathathri-S" }, { "family_name": "Zheng", "given_name": "Stephan", "clpid": "Zheng-Stephan" }, { "family_name": "Gao", "given_name": "Sicun", "clpid": "Gao-Sicun" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" } ], "abstract": "Neural networks are highly sensitive to adversarial examples, which cause large output deviations with only small input perturbations. However, little is known quantitatively about the distribution and prevalence of such adversarial examples. To address this issue, we propose a rigorous search method that provably finds the smallest possible adversarial example. The key benefit of our method is that it gives precise quantitative insight into the distribution of adversarial examples, and guarantees the absence of adversarial examples if they are not found. The primary idea is to consider the nonlinearity exhibited by the network in a small region of the input space, and search exhaustively for adversarial examples in that region. We show that the frequency of adversarial examples and robustness of neural networks is up to twice as large as reported in previous works that use empirical adversarial attacks. In addition, we provide an approach to approximate the nonlinear behavior of neural networks, that makes our search method computationally feasible.", "publisher": "Caltech Library", "publication_date": "2017-11-30" }, { "id": "authors:qp5x4-psg54", "collection": "authors", "collection_id": "qp5x4-psg54", "cite_using_url": "https://resolver.caltech.edu/CaltechAUTHORS:20190709-092100972", "type": "conference_item", "title": "Model-based control of cavity oscillations. II - System identification and analysis", "author": [ { "family_name": "Rowley", "given_name": "Clarence W.", "orcid": "0000-0002-9099-5739", "clpid": "Rowley-C-W" }, { "family_name": "Williams", "given_name": "David R.", "clpid": "Williams-D-R" }, { "family_name": "Colonius", "given_name": "Tim", "orcid": "0000-0003-0326-3909", "clpid": "Colonius-T" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "MacMartin", "given_name": "Douglas G.", "orcid": "0000-0003-1987-9417", "clpid": "MacMartin-D-G" }, { "family_name": "Fabris", "given_name": "Drazin", "clpid": "Fabris-Drazin" } ], "abstract": "Experiments using active control to reduce oscillations in the flow past a rectangular cavity have uncovered surprising phenomena: in the controlled system, often new frequencies of oscillation appear, and often the main frequency of oscillation is split into two sideband frequencies. The goal of this paper is to explain these effects using physics-based models, and to use these ideas to guide control design.\n\nWe present a linear model for the cavity flow, based on the physical mechanisms of the familiar Rossiter model. Experimental data indicates that under many operating conditions, the oscillations are not self-sustained, but in fact are caused by amplification of external disturbances. We present some experimental results demonstrating the peak-splitting phenomena mentioned above, use the physics-based model to study the phenomena, and discuss fundamental performance limitations which limit the achievable performance of any control scheme.", "doi": "10.2514/6.2002-972", "publisher": "American Institute of Aeronautics and Astronautics", "publication_date": "2002-08-22" }, { "id": "authors:r6jj4-0v905", "collection": "authors", "collection_id": "r6jj4-0v905", "cite_using_url": "https://resolver.caltech.edu/CaltechAUTHORS:20190718-165126408", "type": "conference_item", "title": "Model-based control of cavity oscillations. I - Experiments", "author": [ { "family_name": "Williams", "given_name": "David R.", "clpid": "Williams-D-R" }, { "family_name": "Rowley", "given_name": "Clarence W.", "orcid": "0000-0002-9099-5739", "clpid": "Rowley-C-W" }, { "family_name": "Colonius", "given_name": "Tim", "orcid": "0000-0003-0326-3909", "clpid": "Colonius-T" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" }, { "family_name": "MacMartin", "given_name": "Douglas G.", "orcid": "0000-0003-1987-9417", "clpid": "MacMartin-D-G" }, { "family_name": "Fabris", "given_name": "Drazin", "clpid": "Fabris-Drazin" }, { "family_name": "Albertson", "given_name": "Julie", "clpid": "Albertson-Julie" } ], "abstract": "An experimental investigation of acoustic mode noise suppression was conducted in a cavity using a digital controller with a linear control algorithm. The control algorithm was based on flow field physics similar to the Rossiter model for acoustic resonance. Details of the controller and results from its implementation are presented in the companion paper by Rowley, et al.\n\nHere the experiments and some details of the flow field development are described, which were done primarily at Mach number 0.34 corresponding to single mode resonance in the cavity. A novel method using feedback control to suppress the resonant mode and open-loop forcing to inject a non-resonant mode was developed for system identification. The results were used to obtain empirical transfer functions of the components of resonance, and measurements of the shear layer growth for use in the design of the control algorithm.", "doi": "10.2514/6.2002-971", "publisher": "American Institute of Aeronautics and Astronautics", "publication_date": "2002-01" }, { "id": "authors:476q0-75d28", "collection": "authors", "collection_id": "476q0-75d28", "cite_using_url": "https://resolver.caltech.edu/CaltechAUTHORS:20190726-104731145", "type": "conference_item", "title": "Dynamical models for control of cavity oscillations", "author": [ { "family_name": "Rowley", "given_name": "Clarence W.", "orcid": "0000-0002-9099-5739", "clpid": "Rowley-C-W" }, { "family_name": "Colonius", "given_name": "Tim", "orcid": "0000-0003-0326-3909", "clpid": "Colonius-T" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" } ], "abstract": "We investigate nonlinear dynamical models for self-sustained oscillations in the flow past a rectangular cavity. The models are based on the method of Proper Orthogonal Decomposition (POD) and Galerkin projection, and we introduce an inner product and formulation of the equations of motion which enables one to use vector-valued POD modes for compressible flows. We obtain models between 3 and 20 states, which accurately describe both the short-time and long-time dynamics. This is a substantial improvement over previous models based on scalar-valued POD modes, which capture the dynamics for short time, but deviate for long time.", "doi": "10.2514/6.2001-2126", "publisher": "American Institute of Aeronautics and Astronautics", "publication_date": "2001-05" }, { "id": "authors:n9cw0-6z661", "collection": "authors", "collection_id": "n9cw0-6z661", "cite_using_url": "https://resolver.caltech.edu/CaltechAUTHORS:20190726-104731320", "type": "conference_item", "title": "POD Based Models of Self-Sustained Oscillations in the Flow Past an Open Cavity", "author": [ { "family_name": "Rowley", "given_name": "Clarence W.", "orcid": "0000-0002-9099-5739", "clpid": "Rowley-C-W" }, { "family_name": "Colonius", "given_name": "Tim", "orcid": "0000-0003-0326-3909", "clpid": "Colonius-T" }, { "family_name": "Murray", "given_name": "Richard M.", "orcid": "0000-0002-5785-7481", "clpid": "Murray-R-M" } ], "abstract": "The goal of this work is to provide accurate dynamical models of oscillations in the flow past a rectangular cavity, for the purpose of bifurcation analysis and control. We have performed an extensive set of direct numerical simulations which provide the data used to derive and evaluate the models. Based on the method of Proper Orthogonal Decomposition (POD) and Galerkin projection, we obtain low-order models (from 6 to 60 states) which capture the dynamics very accurately over a few periods of oscillation, but deviate for long time.", "doi": "10.2514/6.2000-1969", "publisher": "American Institute of Aeronautics and Astronautics", "publication_date": "2000-06" } ]