Compute-intensive Models Research Articles

Bayesian monitoring of substructures under unknown interface assumption

Structural Health Monitoring (SHM) enables assessing in-service structures’ performance by localizing structural anomaly instances immediately after their occurrence. Typical SHM approaches monitor the entire structural spatial domain aggravating the required density and cost of instrumentation. Further, with model-based approaches, the entire structural domain is needed to be defined with high dimensional, compute-intensive models rendering the SHM approaches ill-posed and slow especially when the instrumentation is limited and system observability is compromised. Moreover, in absence of high-fidelity models, oversimplification and subsequent model inaccuracies may lead to inaccurate estimation and possibly false alarms even if a subdomain is modeled inaccurately, e.g. support boundaries. To mitigate such issues, stand-alone monitoring focusing only on a subdomain of interest may be a computationally cheaper and prompt approach while being substantially robust to false alarms. Typically, such stand-alone substructure monitoring approaches demand extensive measurement of the interface, which can be a challenge in real-life applications. This paper presents a novel filtering-based online time domain approach for estimating substructure parameters without the need to measure or estimate the substructure interfaces. The proposed component-wise estimation is stand-alone so that the health estimation of the complete structural domain can be undertaken in parallel and later coupled through post-processing. The requirement of the interface measurement has been alleviated by employing an output injection approach. The proposal has been validated on a numerical beam structure subjected to arbitrary forces and subsequently, the sensitivity against noise and damage severity of the proposal has been investigated. Finally, the proposal is validated on a real beam to illustrate its real-life applicability and significance.

Network offloading policies for cloud robotics: a learning-based approach

Today’s robotic systems are increasingly turning to computationally expensive models such as deep neural networks (DNNs) for tasks like localization, perception, planning, and object detection. However, resource-constrained robots, like low-power drones, often have insufficient on-board compute resources or power reserves to scalably run the most accurate, state-of-the art neural network compute models. Cloud robotics allows mobile robots the benefit of offloading compute to centralized servers if they are uncertain locally or want to run more accurate, compute-intensive models. However, cloud robotics comes with a key, often understated cost: communicating with the cloud over congested wireless networks may result in latency or loss of data. In fact, sending high data-rate video or LIDAR from multiple robots over congested networks can lead to prohibitive delay for real-time applications, which we measure experimentally. In this paper, we formulate a novel Robot Offloading Problem—how and when should robots offload sensing tasks, especially if they are uncertain, to improve accuracy while minimizing the cost of cloud communication? We formulate offloading as a sequential decision making problem for robots, and propose a solution using deep reinforcement learning. In both simulations and hardware experiments using state-of-the art vision DNNs, our offloading strategy improves vision task performance by between 1.3 and 2.3 $$\times $$ of benchmark offloading strategies, allowing robots the potential to significantly transcend their on-board sensing accuracy but with limited cost of cloud communication.

Surrogate-based black-box optimisation via domain exploration and smart placement

In the era of digital twins, the need for accurately mimicking the reality has given rise to complex, black-box, compute-intensive models that are vital for simulating, analysing, and optimising physicochemical systems. In this work, we propose a novel surrogate-assisted approach for black-box optimisation, which uses efficient domain exploration and smart adaptive sample placement to escape local valleys (traps) and obtain a global minimum efficiently. Our iterative algorithm comprises two stages. The first stage constructs sub-regions based on Delaunay triangulations and selects the best for exploration. The second stage adds a new sample point to the best sub-region via optimisation. The two stages together balance domain exploration versus exploitation. The algorithmic framework is illustrated using the six-hump camel back function. An extensive numerical evaluation using twenty test functions (up to six variables) shows that the proposed algorithm exhibits superior performance against seven well-known commercial global optimisation algorithms including a surrogate-based approach.

Detecting Genetic Risk Factors for Alzheimer's Disease in Whole Genome Sequence Data via Lasso Screening.

Genetic factors play a key role in Alzheimer's disease (AD). The Alzheimer's Disease Neuroimaging Initiative (ADNI) whole genome sequence (WGS) data offers new power to investigate mechanisms of AD by combining entire genome sequences with neuroimaging and clinical data. Here we explore the ADNI WGS SNP (single nucleotide polymorphism) data in depth and extract approximately six million valid SNP features. We investigate imaging genetics associations using Lasso regression-a widely used sparse learning technique. To solve the large-scale Lasso problem more efficiently, we employ a highly efficient screening rule for Lasso-called dual polytope projections (DPP)-to remove irrelevant features from the optimization problem. Experiments demonstrate that the DPP can effectively identify irrelevant features and leads to a 400× speedup. This allows us for the first time to run the compute-intensive model selection procedure called stability selection to rank SNPs that may affect the brain and AD risk.