More Like this

1. Interval Reachability of Nonlinear Dynamical Systems with Neural Network Controllers

   Jafarpour, Saber; Harapanahalli, Akash; Coogan, Samuel (June 2023, 5th Annual Learning for Dynamics & Control Conference)

   N. Matni, M. Morari (Ed.)

   This paper proposes a computationally efficient framework, based on interval analysis, for rigorous verification of nonlinear continuous-time dynamical systems with neural network controllers. Given a neural network, we use an existing verification algorithm to construct inclusion functions for its input-output behavior. Inspired by mixed monotone theory, we embed the closed-loop dynamics into a larger system using an inclusion function of the neural network and a decomposition function of the open-loop system. This embedding provides a scalable approach for safety analysis of the neural control loop while preserving the nonlinear structure of the system. We show that one can efficiently compute hyper-rectangular over-approximations of the reachable sets using a single trajectory of the embedding system. We design an algorithm to leverage this computational advantage through partitioning strategies, improving our reachable set estimates while balancing its runtime with tunable parameters. We demonstrate the performance of this algorithm through two case studies. First, we demonstrate this method’s strength in complex nonlinear environments. Then, we show that our approach matches the performance of the state-of-the art verification algorithm for linear discretized systems. 

   more » « less
2. Clustering time series with nonlinear dynamics: A Bayesian non-parametric and particle-based approach

   Lin, A.; Zhang, Y.; Heng, J.; Allsop, S.; Tye, K.; Jacob, P.; Ba, D. (April 2019, AISTATS)

   We propose a statistical framework for clustering multiple time series that exhibit nonlinear dynamics into an a-priori-unknown number of sub-groups that each comprise time series with similar dynamics. Our motivation comes from neuroscience where an important problem is to identify, within a large assembly of neurons, sub-groups that respond similarly to a stimulus or contingency. In the neural setting, conditioned on cluster membership and the parameters governing the dynamics, time series within a cluster are assumed independent and generated according to a nonlinear binomial state-space model. We derive a Metropolis-within-Gibbs algorithm for full Bayesian inference that alternates between sampling of cluster membership and sampling of parameters of interest. The Metropolis step is a PMMH iteration that requires an unbiased, low variance estimate of the likelihood function of a nonlinear state- space model. We leverage recent results on controlled sequential Monte Carlo to estimate likelihood functions more efficiently compared to the bootstrap particle filter. We apply the framework to time series acquired from the prefrontal cortex of mice in an experiment designed to characterize the neural underpinnings of fear. 

   more » « less
3. Optimal Scheduling for Unmanned Aerial Vehicle Networks with Flow-Level Dynamics

   https://doi.org/10.1109/TMC.2019.2952848  

   Kong, Xiangqi; Lu, Ning; Li, Bin (November 2019, IEEE Transactions on Mobile Computing)

   Unmanned Aerial Vehicle (UAV) Networks have recently attracted great attention as being able to provide convenient and fast wireless connections. One central question is how to allocate a limited number of UAVs to provide wireless services across a large number of regions, where each region has dynamic arriving flows and flows depart from the system once they receive the desired amount of service (referred to as the flow-level dynamic model). In this paper, we propose a MaxWeight-type scheduling algorithm taking into account sharp flow-level dynamics that efficiently redirect UAVs across a large number of regions. However, in our considered model, each flow experiences an independent fading channel and will immediately leave the system once it completes its service, which makes its evolution quite different from the traditional queueing model for wireless networks. This poses significant challenges in our performance analysis. Nevertheless, we incorporate sharp flow-dynamic into the Lyapunov-drift analysis framework, and successfully establish both throughput and heavy-traffic optimality of the proposed algorithm. Extensive simulations are performed to validate the effectiveness of our proposed algorithm. 

   more » « less
4. Satisfaction-aware Data Offloading in Surveillance Systems

   https://doi.org/10.1145/3349625.3355437  

   Apostolopoulos, Pavlos Athanasios; Torres, Marcos; Tsiropoulou, Eirini Eleni (January 2019, CHANTS'19: Proceedings of the 14th Workshop on Challenged Networks)

   null (Ed.)

   In this paper, we exploit the capabilities of Fully Autonomous Aerial Systems' (FAAS) and the Mobile Edge Computing (MEC) to introduce a novel data offloading framework and support the energy and time efficient video processing in surveillance systems based on game theory in satisfaction form. A surveillance system is introduced consisting of Areas of Interest (AoIs), where a MEC server is associated with each AoI, and a FAAS is flying above the AoIs to collectively support the IP cameras' computing demands. Each IP camera adopts a utility function capturing its Quality of Service (QoS) considering the experienced time and energy overhead to offload and process its data either remotely or locally. A non-cooperative game among the cameras is formulated to determine the amount of offloading data to the MEC server and/or the FAAS. The novel concept of Satisfaction Equilibrium (SE) is introduced where the IP cameras satisfy their minimum QoS prerequisites instead of maximizing their performance by wasting additional system resources. A distributed learning algorithm determines the IP cameras' stable data offloading, while a reinforcement learning algorithm determines the FAAS's movement among the AoIs exploiting the accuracy, timeliness, and certainty of the collected data by the IP cameras per AoI. The performance evaluation of the proposed framework is achieved via modeling and simulation. 

   more » « less
5. A Human-Centered Power Conservation Framework Based on Reverse Auction Theory and Machine Learning

   https://doi.org/10.1145/3656348  

   Casella, Enrico; Silvestri, Simone; Baker, D A; Das, Sajal K (July 2024, ACM Transactions on Cyber-Physical Systems)

   Extreme outside temperatures resulting from heat waves, winter storms, and similar weather-related events trigger the Heating Ventilation and Air Conditioning (HVAC) systems, resulting in challenging, and potentially catastrophic, peak loads. As a consequence, such extreme outside temperatures put a strain on power grids and may thus lead to blackouts. To avoid the financial and personal repercussions of peak loads, demand response and power conservation represent promising solutions. Despite numerous efforts, it has been shown that the current state-of-the-art fails to consider (1) the complexity of human behavior when interacting with power conservation systems and (2) realistic home-level power dynamics. As a consequence, this leads to approaches that are (1) ineffective due to poor long-term user engagement and (2) too abstract to be used in real-world settings. In this article, we propose an auction theory-based power conservation framework for HVAC designed to address such individual human component through a three-fold approach:personalized preferencesof power conservation,models of realistic user behavior, andrealistic home-level power dynamics. In our framework, the System Operator sends Load Serving Entities (LSEs) the required power saving to tackle peak loads at the residential distribution feeder. Each LSE then prompts its users to providebids, i.e.,personalized preferencesof thermostat temperature adjustments, along with corresponding financial compensations. We employmodels of realistic user behaviorby means of online surveys to gather user bids and evaluate user interaction with such system.Realistic home-level power dynamicsare implemented by our machine learning-based Power Saving Predictions (PSP) algorithm, calculating the individual power savings in each user’s home resulting from such bids. A machine learning-based PSPs algorithm is executed by the users’ Smart Energy Management System (SEMS). PSP translates temperature adjustments into the corresponding power savings. Then, the SEMS sends bids back to the LSE, which selects the auction winners through an optimization problem called POwer Conservation Optimization (POCO). We prove that POCO is NP-hard, and thus provide two approaches to solve this problem. One approach is an optimal pseudo-polynomial algorithm called DYnamic programming Power Saving (DYPS), while the second is a heuristic polynomial time algorithm called Greedy Ranking AllocatioN (GRAN). EnergyPlus, the high-fidelity and gold-standard energy simulator funded by the U.S. Department of Energy, was used to validate our experiments, as well as to collect data to train PSP. We further evaluate the results of the auctions across several scenarios, showing that, as expected, DYPS finds the optimal solution, while GRAN outperforms recent state-of-the-art approaches. 

   more » « less