Accountability and Evaluation of AI Algorithms

Permanent URI for this collection

https://hdl.handle.net/10125/63600

Browse

Recent Submissions

Now showing 1 - 3 of 3
  • Item
    Reliability of Training Data Sets for ML Classifiers: A Lesson Learned from Mechanical Engineering
    ( 2020-01-07) Juric, Radmila ; Danilchanka, Natallia ; Mousavi, Mehdi Gebreil
    The popularity of learning and predictive technologies, across many problem domains, is unprecedented and it is often underpinned with the fact that we efficiently compute with vast amounts of data and data types, and thus should be able to resolve problems, which we could not in the past. This view is particularly common among scientists who believe that the excessive amount of data, we generate in real life, is ideal for performing predictions and training algorithms. However, the truth might be quite different. The paper illustrates the process of preparing a training data set for an ML classifier, which should predict certain conditions in mechanical engineering. It was not the case that it was difficult to define and choose classifiers, in order to secure safe predictions. It was our inability to create a safe, reliable and trustworthy training data set, from scientifically proven experiments, which created the problem. This places serious doubts on the way we use learning and predictive technologies today. It remains debatable what the next step should be. However, if in ML algorithms, and classifiers in particular, the semantic which is built-in data sets, influences classifier’s definition, it would be very difficult to evaluate and rely on them, before we understand data semantics fully. In other words, we still do not know how the semantic, sometimes hidden in a data set, can adversely affect algorithms trained by them.
  • Item
    A Comparison Between a Two-Feedback Control Loop and a Reinforcement Learning Algorithm for Compliant Low-Cost Series Elastic Actuators
    ( 2020-01-07) Sanfilippo, Filippo ; Tuan Minh, Hua ; Bos, Steven
    Highly-compliant elastic actuators have become progressively prominent over the last years for a variety of robotic applications. With remarkable shock tolerance, elastic actuators are appropriate for robots operating in unstructured environments. In accordance with this trend, a novel elastic actuator was recently designed by our research group for Serpens, a low-cost, open-source and highly-compliant multi-purpose modular snake robot. To control the newly designed elastic actuators of Serpens, a two-feedback loops position control algorithm was proposed. The inner controller loop is implemented as a model reference adaptive controller (MRAC), while the outer control loop adopts a fuzzy proportional-integral controller (FPIC). The performance of the presented control scheme was demonstrated through simulations. However, the efficiency of the proposed controller is dependent on the initial values of the parameters of the MRAC controller as well as on the effort required for a human to manually construct fuzzy rules. An alternative solution to the problem might consist of using methods that do not assume a priori knowledge: a solution that derives its properties from a machine learning procedure. In this way, the controller would be able to automatically learn the properties of the elastic actuator to be controlled. In this work, a novel controller for the proposed elastic actuator is presented based on the use of an artificial neural network (ANN) that is trained with reinforcement learning. The newly designed control algorithm is extensively compared with the former approach. Simulation results are presented for both methods. The authors seek to achieve a fair, non-biased, risk-aware and trustworthy comparison.
  • Item