This paper proposes the first known approaches to locating an athletics track from GPS data of a running workout. This is done by mapping the GPS points onto an image and applying Computer Vision methods, including: performing keypoint detection using a Keypoint Region-Based Convolutional Neural Network (R-CNN) model with a Residual Neural Network backbone; and applying Convolutional Kernels over the image. The Keypoint R-CNN model achieved a Root Mean Square Error (RMSE) of 1.29 on synthetic data, which is lower than the error of other keypoint detection models of 2.23, but performed poorly on real data. The Convolutional Method achieved an RMSE of 4.20. The methods and results of this paper can be used in future applications for error correction and feature extraction of GPS data for running workout analysis.

Transfer learning is the ability to transfer knowledge from one context to another. This paper investigates, for the first time, the possibility of transfer learning on Monte Carlo Tree Search (MCTS). We use distributed parallel MCTS to construct game-playing models for one variant of chess and then transfer these models to another domain to play a different variant of chess. We also introduce a novel pruning mechanism for MCTS to cope with memory requirements of large game trees. Our results show that positive transfer learning in MCTS is possible, and in our case, highly successful. Transferred models significantly outperform models that have to be trained from scratch.

Developing and maintaining constructive engagement is a crucial challenge in learning by watching videos. AVW-Space is an online video-based learning platform which enhances student engagement via note-taking and personalised support. Previous studies with AVW-Space show that students who write comments, especially high-quality comments, learn more. The goal of the study reported in this paper is to encourage students to write better-quality comments. After automating the assessment of comment quality using machine learning approaches, we developed Quality nudges which encourage students to write better comments. We conducted a study in a first-year engineering course to analyse the learning effects of the Quality nudges. The results show that Quality nudges enhanced constructive engagement and learning. The contribution of this research is in proposing methodology for increasing the quality of student comments in video-based learning.

Active Video Watching (AVW-Space) is an online platform for video-based learning which supports engagement via note-taking and personalized nudges. In this paper, we focus on the quality of the comments students write. We propose two schemes for assessing the quality of comments. Then, we evaluate these schemes by computing the inter-coder agreement. We also evaluate various machine learning classifiers to automate the assessment of comments. The selected cost-sensitive classifier shows that the quality of comments can be assessed with high weighted-F1 scores. This study contributes to the automation of comment quality assessment and the development of personalized educational support for engagement in video-based learning through commenting.

Deep reinforcement learning, the combination of deep learning and reinforcement learning, has enabled the training of agents that can solve complex tasks from visual inputs. However, these methods often require prohibitive amounts of computation to obtain successful results. To improve learning efficiency, there has been a renewed focus on separating state representation and policy learning. In this paper, we investigate the quality of state representations learned by different types of autoencoders, a popular class of neural networks used for representation learning. We assess not only the quality of the representations learned by undercomplete, variational, and disentangled variational autoencoders, but also how the quality of the learned representations is affected by changes in representation size. To accomplish this, we also present a new method for evaluating learned state representations for Atari games using the Atari Annotated RAM Interface. Our findings highlight differences in the quality of state representations learned by different types of autoencoders and their robustness to reduction in representation size. Our results also demonstrate the advantage of using more sophisticated evaluation methods over assessing reconstruction quality.

Recent success in solving hard reinforcement learning problems can be partly credited to the use of deep neural networks, which can extract high-level features and learn compact state representations from high-dimensional inputs, such as images. However, the large networks required to learn both state representation and policy using this approach limit the effectiveness and benefits of neuroevolution methods that have proven effective at solving simpler problems in the past. One potential solution to this problem is to separate state representation and policy learning and only apply neuroevolution to the latter. We extend research following this approach by evolving small policy networks for Atari games using NEAT, that learn from compact state representations provided by the recently released Atari Annotated RAM Interface (Atari ARI). Our results show that it is possible to evolve agents that exceed expert human performance using these compact state representations, and that, for some games, successful policy networks can be evolved that contain only a few or even no hidden nodes.

In this paper, we prove the existence of feature bundles in some classification problems. These are a set of features that while in their entirety are relevant to the target variable, any strict subset of them is completely independent of the target. Any machine learning algorithm applied to a strict subset of a feature bundle cannot produce a model that performs better than a feature-less model that always predicts the majority class. We demonstrate and discuss the effect of these feature bundles on the performance of tree-based classification learning and feature selection algorithms.

The overarching goal of our project is to design effective learning activities for PyKinetic, a smartphone Python tutor. In this paper, we present a study using a variant of Parsons problems we designed for PyKinetic. Parsons problems contain randomized code which needs to be re-ordered to produce the desired effect. In our variant of Parsons problems, students were asked to complete the missing part(s) of some lines of code (LOCs), and rearrange the LOCs to match the problem description. In addition, we added menu-based Self-Explanation (SE) prompts. Students were asked to self-explain concepts related to incomplete LOCs they solved. Our hypotheses were: (H1) PyKinetic would be successful in supporting learning; (H2) menu-based SE prompts would result in further learning benefits; (H3) students with low prior knowledge (LP) would learn more from our Parsons problems in comparison to those with high prior knowledge (HP). We found that the participants' scores on the post-test improved, thus showing evidence of learning in PyKinetic. The experimental group participants, who had SE prompts, showed improved learning in comparison to the control group. Further analyses revealed that LP students improved more than HP students and the improvement is even more pronounced for LP learners who self-explained. The contributions of our work are a) pedagogically-guided design of Parsons problems with SE prompts used on smartphones, b) showing that our Parsons problems are effective in supporting learning and c) our Parsons problems with SE prompts are especially effective for students with low prior knowledge.

Feature selection methods are generally divided into three categories: filter, wrapper and embedded approaches. In terms of learning performance, the filter approach is typically inferior compared to the other two because it does not use the target learning algorithm. The embedded and wrapper approaches are both considered high-performing. In this paper we compare the embedded and the wrapper approaches in the context of Support Vector Machines (SVMs). In the wrapper category, we compare well-known algorithms such as Genetic Algorithm (GA), Forward and Backward selection, and a new binary Particle Swarm Optimization (PSO) algorithm. For an embedded approach we devise a new heuristic algorithm based on Multiple Kernel Learning.

Binary PSO algorithms are extensions of the PSO algorithm that enjoy some of the social intelligence properties of the original algorithm. The intensive local search ability is one of the most important characteristics of PSO. In this paper, we argue that, when evaluating binary PSO algorithms against common real-value benchmark problems—a common practice in the literature—the representation of the search space can have a significant effect on the results. For this purpose we propose the use of reflected binary code, which is a minimal change ordering representation for mapping a binary genotype space to a real phenotype space, while preserving the notion of locality in the phenotype space.

Mobile devices are increasingly being utilized for learning due to their unique features including portability for providing ubiquitous experiences. In this paper, we present PyKinetic, a mobile tutor we developed for Python programming, aimed to serve as a supplement to traditional courses. The overarching goal of our work is to design coding activities that maximize learning. As we work towards our goal, we first focus on the learning effectiveness of the activities within PyKinetic, rather than evaluating the effectiveness of PyKinetic as a supplement resource for an introductory programming course. The version of PyKinetic (PyKinetic_DbgOut) used in the study contains five types of learning activities aimed at supporting debugging, code-tracing, and code writing skills. We evaluated PyKinetic in a controlled lab study with quantitative and qualitative results to address the following research questions: (R1) Is the combination of coding activities effective for learning programming? (R2) How do the activities affect the skills of students with lower prior knowledge (novices) compared to those who had higher prior knowledge (advanced)? (R3) How can we improve the usability of PyKinetic? Results revealed that PyKinetic_DbgOut was more beneficial for advanced students. Furthermore, we found how coding skills are interrelated differently for novices compared to advanced learners. Lastly, we acquired sufficient feedback from the participants to improve the tutor.

We present our study on PyKinetic with various activities to target several skills: code tracing, debugging, and code writing. Half of the participants (control group) received the problems in a fixed order, while for the other half (experimental group) problems were selected adaptively, based on their performance. In a previous paper, we discussed the general findings from the study. In this paper we present further analyses and focus on differences between low performing students and students with higher pre-existing knowledge. We hypothesized that: (H1) novices will benefit more than advanced students, and (H2) advanced students in the experimental group will benefit more than those in the control group. The results confirmed H1 and revealed that this version of PyKinetic was more beneficial for novice learners. Moreover, novices showed evidence of learning multiple skills: code writing, debugging and code tracing. However, we did not have enough evidence for hypothesis H2.

Image segmentation is still an open problem in image processing due to the demand of efficient and effective segmentation algorithm. Researchers face many challenges in developing a fully automatic segmentation algorithm for complex scenario. This situation motivated us to explore another concept that can represent the complex scenario well. This paper presents an application of the neutrosophic logic as the generalization of fuzzy logic for extracting visible brushstroke regions in paintings by Vincent van Gogh. Research suggests that neutrosophic logic can represent ambiguity and uncertainty in a better way than fuzzy logic, thus can be used to represent noise and uncertain information that exist in an image. In this work, we observe some digital paintings of Van Gogh by filtering out the existing noise using circular filter and converting them into the neutrosophic domain. The combination of those two methods result in clean, magnified visible brushstroke edges which define important brush region clearly.

Apple harvest is a time-consuming process which requires intensive skilled labor. Precision agriculture and automation solutions are facing several challenges such as entangled tasks, complex dynamics, challenging terrains, and cost-effectiveness. This paper introduces a hybrid control multi-agent cooperative system to incorporate pickers' manual skills and flexibility with robots precision, and navigation ability to automate bins transportation in order to optimize harvest time and cost. This paper presents a multi-agent cooperative system and a simulation of a tractor to transport fixed fruit bins versus a simulation of a robot agent employed by the system as an autonomous bin.

We have developed PyKinetic, a mobile Python tutor for novices. We present our study on PyKinetic with various activities to target several skills: code tracing, debugging, code understanding and code writing. We compared a version with a fixed sequence of learning activities to an adaptive version, containing the same activities but with personalized problem selection. We had two hypotheses: (H1) the combination of activities is effective for learning, and (H2) the adaptive problem selection is beneficial. The results show that PyKinetic is effective for learning, and the adaptive version provides additional benefits for learners.

PyKinetic is a mobile tutor for Python, which offers Parsons problems with incomplete lines of code (LOCs). This paper reports the results of a study in which we investigated the effect of menu-based self-explanation (SE) prompts. Students were asked to self-explain concepts related to incomplete LOCs they have solved. The goals of the study were (1) to investigate whether students are learning with PyKinetic and (2) to determine the effect of SE prompts. The scores of participants have significantly improved from the pre-test to the post-test. There was also a significant difference on the post-test scores of participants from the experimental group compared to the control group. In future work, we aim to add other activities to PyKinetic, and introduce a student model and a pedagogical model for an adaptive version of PyKinetic.

This paper presents an algorithm for brush stroke characterization to capture the style of Van Gogh through extracting the shape and texture features. Those features were then tested for their representability using machine learning techniques, proving 89.53% of accuracy in differentiating Van Gogh from his contemporary. This paper also outlines the modelling of a series of digital brush strokes using a coordinated particle system. The tested features are transformed into parameters for the brush stroke model to render an input image in the style of Van Gogh.

Kernel functions have an important role in the performance of Support Vector Machines (SVMs), since they form the geometry of the feature space. Manual designing of kernel functions is an expensive task and requires domain-specific knowledge. In this article, we propose a new method to automatically construct kernel functions and select optimal subsets of features. We achieve this by combining primitive kernels and subsets of features using Genetic Programming (GP). Our experiments show that the proposed method drastically improves the prediction accuracy of SVMs.

Recent developments of unconventional hardware using memristors and atomic switch networks has led to renewed interest in hardware neuromorphic solutions. Most hardware models rely upon a reservoir neural network as the basis of any learning, but the distinct differences between software implementations and hardware reality mean what we take for granted in traditional software reservoirs — such as cycles, loops, infinite energy, and discrete time — may be severely limited or unavailable in hardware, raising questions about how a hardware implementation would perform and how to potentially overcome these limitations. Proposed hardware additions, such as an echoer or an input delay mechanism, address some of these limitations.