Interactive Learning with Corrective Feedback for Policies Based on Deep Neural Networks

With the Internet of Things paradigm, the data generated by the rapidly increasing number of connected devices lead to new possibilities, such as using machine learning for activity recognition in smart environments. However, it also introduces several challenges. The sensors of different devices might be of different types, making the fusion of data non-trivial. Moreover, the devices are often mobile, resulting in that data from a particular sensor is not always available, i.e. there is a need to handle data from a dynamic set of sensors. From a machine learning perspective, the data from the sensors arrives in a streaming fashion, i.e., online learning, as compared to many learning problems where a static dataset is assumed. Machine learning is in many cases a good approach for classification problems, but the performance is often linked to the quality of the data. Having a good data set to train a model can be an issue in general, due to the often costly process of annotating the data. With dynamic and heterogeneous data, annotation can be even more problematic, because of the ever-changing environment. This means that there might not be any, or a very small amount of, annotated data to train the model on at the start of learning, often referred to as the cold start problem. To be able to handle these issues, adaptive systems are needed. With adaptive we mean that the model is not static over time, but is updated if there for instance is a change in the environment. By including human-in-the-loop during the learning process, which we refer to as interactive machine learning, the input from users can be utilized to build the model. The type of input used is typically annotations of the data, i.e. user input in the form of correctly labelled data points. Generally, it is assumed that the user always provides correct labels in accordance with the chosen interactive learning strategy. In many real-world applications these assumptions are not realistic however, as users might provide incorrect labels or not provide labels at all in line with the chosen strategy. In this thesis we explore which interactive learning strategies are possible in the given scenario and how they affect performance, as well as the effect of machine learning algorithms on performance. We also study how a user who is not always reliable, i.e. that does not always provide a correct label when expected to, can affect performance. We propose a taxonomy of interactive online machine learning strategies and test how the different strategies affect performance through experiments on multiple datasets. The findings show that the overall best performing interactive learning strategy is one where the user provides labels when previous estimations have been incorrect, but that the best performing machine learning algorithm depends on the problem scenario. The experiments also show that a decreased reliability of the user leads to decreased performance, especially when there is a limited amount of labelled data.

Learning ability is the basic characteristic of human intelligence. The July 1, 2005 issue of Science published a list of 125 important questions in science. Among them, the question 94 “What are the limits of learning by machines?”. The annotation “Computers can already beat the world’s best chess players, and they have a wealth of information on the Web to draw on. But abstract reasoning is still beyond any machine”. In recent artificial intelligence has made great progresses. In 1997, the rise of the man-machine war, IBM Supercomputer Deep Blue defeated the chess master Garry Kasparov. On February 14, 2011, IBM’s Watson supercomputer won a practice round against Jeopardy champions Ken Jennings and Brad Rutter. In March 2016, Google DeepMind’s AlphaGo sealed a 4-1 victory over a South Korean Go grandmaster Lee Se-dol. This paper focuses on the machine learning methods of AlphaGo, including reinforcement learning, deep learning, deep reinforcement learning, analysis of the existing problems and the latest research progress. Deep reinforcement learning is the combination of deep learning and reinforcement learning, which can realize the learning algorithm from the perception to action. Simply said, this is the same as human behavior, input sensing information such as vision, and then, direct output action through the deep neural network. Deep reinforcement learning has the potential to learn a variety of skills for the robot to achieve full autonomy. Even though reinforcement learning is practiced successfully, but feature states need to manually set, for complex scene is a difficult thing, especially easy to cause the dimension disaster, and expression is not good. In 2010, Sascha Lange and Martin Riedmiller proposed deep auto-encoder neural networks in reinforcement learning to extract feature, which is used to control the visual correlation. In 2013, DeepMind proposed deep Q-network (DQN) in NIPS 2013, using convolution neural network to extract features, and then applied in reinforcement learning. They continue to improve and published an improved version of DQN on Nature in 2015, which has aroused widespread concern. In order to break through the limits of learning by machines, cognitive machine learning is proposed, which is the combination of machine learning and brain cognition, so that the machine intelligence is constantly evolving, and gradually reaches the human level of artificial intelligence. A cognitive model entitled Consciousness And Memory (CAM) is proposed by author, which consists of memory, consciousness, high-level cognitive functions, perception and motor. High-level cognitive functions of the brain include learning, language, thinking, decision making, emotion, and so on. Learning is a course to accept the stimulus through the nervous system and obtain new behavior, habits and accumulation experience. According to the current research progress of brain science and cognitive science, cognitive machine learning may be interested in learning emergence, procedural memory knowledge learning, learning evolution and so on. For intelligence, so-called evolution is refers to the learning of learning and the structure also follows the change. It is important to record the learning result by structure changing and improve the learning method.

Deep reinforcement learning (DRL) has superior autonomous decision-making capabilities, combining deep learning and reinforcement learning (RL). Unlike DRL employs deep neural networks (DNNs), broad RL (BRL) adopts the broad learning system (BLS) that is established with flat networks to generate the strategy. This article proposes the multiagent adaptive broad-DRL (ABDRL) approach for traffic light control (TLC), which combines the broad network with the deep network structure. Specifically, the structure of ABDRL first expands in the form of flatted broad networks. Then, the feature representation module that contains DNNs is employed to extract the critical traffic information. In addition, experiences sampled randomly by the experience replay mechanism cannot reflect the current training status of the agent effectively. In order to alleviate the impacts caused by random sampling, the forgetful experience mechanism (FEM) is incorporated into ABDRL. The FEM enables the agent to discriminate the importance of experiences stored in the experience reply buffer to improve robustness and adaptability. We validate the effectiveness of ABDRL in TLC, and the results illustrate the optimality and robustness of ABDRL over the state-of-the-art multiagent DRL (MADRL) algorithms.

Deep reinforcement learning (DRL) has proven capable of superhuman performance on many complex tasks. To achieve this success, DRL algorithms train a decision-making agent to select the actions that maximize some long-term performance measure. In many consequential real-world domains, however, optimal performance is not enough to justify an algorithm’s use—for example, sometimes a system’s robustness, stability, or safety must be rigorously ensured. Thus, methods for verifying DRL systems have emerged. These algorithms can guarantee a system’s properties over an infinite set of inputs, but the task is not trivial. DRL relies on deep neural networks (DNNs). DNNs are often referred to as “black boxes” because examining their respective structures does not elucidate their decision-making processes. Moreover, the sequential nature of the problems DRL is used to solve promotes significant scalability challenges. Finally, because DRL environments are often stochastic, verification methods must account for probabilistic behavior. To address these complications, a new subfield has emerged. In this survey, we establish the foundations of DRL and DRL verification, define a taxonomy for DRL verification methods, describe approaches for dealing with stochasticity, characterize considerations related to writing specifications, enumerate common testing tasks/environments, and detail opportunities for future research.

In this paper, we highlight three issues that limit performance of machine learning on biomedical images, and tackle them through 3 case studies: 1) Interactive Machine Learning (IML): we show how IML can drastically improve exploration time and quality of direct volume rendering. 2) transfer learning: we show how transfer learning along with intelligent pre-processing can result in better Alzheimer's diagnosis using a much smaller training set 3) data imbalance: we show how our novel focal Tversky loss function can provide better segmentation results taking into account the imbalanced nature of segmentation datasets. The case studies are accompanied by in-depth analytical discussion of results with possible future directions.

Reinforcement learning is a mathematical framework for agents to interact intelligently with their environment. Unlike supervised learning, where a system learns with the help of labeled data, reinforcement learning agents learn how to act by trial and error only receiving a reward signal from their environments. A field where reinforcement learning has been prominently successful is robotics [3]. However, real-world control problems are also particularly challenging because of the noise and high- dimensionality of input data (e.g., visual input). In recent years, in the field of supervised learning, deep neural networks have been successfully used to extract meaning from this kind of data. Building on these advances, deep reinforcement learning was used to solve complex problems like Atari games and Go. Mnih et al. [1] built a system with fixed hyper parameters able to learn to play 49 different Atari games only from raw pixel inputs. However, in order to apply the same methods to real-world control problems, deep reinforcement learning has to be able to deal with continuous action spaces. Discretizing continuous action spaces would scale poorly, since the number of discrete actions grows exponentially with the dimensionality of the action. Furthermore, having a parametrized policy can be advantageous because it can generalize in the action space. Therefore with this thesis we study state-of-the-art deep reinforcement learning algorithm, Deep Deterministic Policy Gradients. We provide a theoretical comparison to other popular methods, an evaluation of its performance, identify its limitations and investigate future directions of research. The remainder of the thesis is organized as follows. We start by introducing the field of interest, machine learning, focusing our attention of deep learning and reinforcement learning. We continue by describing in details the two main algorithms, core of this study, namely Deep Q-Network (DQN) and Deep Deterministic Policy Gradients (DDPG). We then provide implementatory details of DDPG and our test environment, followed by a description of benchmark test cases. Finally, we discuss the results of our evaluation, identifying limitations of the current approach and proposing future avenues of research.

Robot learning problems are limited by physical constraints, which make learning successful policies for complex motor skills on real systems unfeasible. Some reinforcement learning methods, like Policy Search, offer stable convergence toward locally optimal solutions, whereas interactive machine learning or learning-from-demonstration methods allow fast transfer of human knowledge to the agents. However, most methods require expert demonstrations. In this work, we propose the use of human corrective advice in the actions domain for learning motor trajectories. Additionally, we combine this human feedback with reward functions in a Policy Search learning scheme. The use of both sources of information speeds up the learning process, since the intuitive knowledge of the human teacher can be easily transferred to the agent, while the Policy Search method with the cost/reward function take over for supervising the process and reducing the influence of occasional wrong human corrections. This interactive approach has been validated for learning movement primitives with simulated arms with several degrees of freedom in reaching via-point movements, and also using real robots in such tasks as “writing characters” and the ball-in-a-cup game. Compared with standard reinforcement learning without human advice, the results show that the proposed method not only converges to higher rewards when learning movement primitives, but also that the learning is sped up by a factor of 4–40 times, depending on the task.

In the domain of continuous control, deep reinforcement learning (DRL) demonstrates promising results. However, the dependence of DRL on deep neural networks (DNNs) results in the demand for extensive data and increased computational cost. To address this issue, a novel hybrid actor-critic reinforcement learning (RL) framework is introduced. The proposed framework integrates the broad learning system (BLS) with DNN, aiming to merge the strengths of both distinct architectural paradigms. Specifically, the critic network employs BLS for rapid value estimation via ridge regression, while the actor network retains the DNN structure to optimize policy gradients. This hybrid design is generalizable and can enhance existing actor-critic algorithms. To demonstrate its versatility, the proposed framework is integrated into three widely used actor-critic algorithms-deep deterministic policy gradient (DDPG), soft actor-critic (SAC), and twin delayed DDPG (TD3), resulting in BLS-augmented variants. The experimental results reveal that all BLS-enhanced versions surpass their original counterparts in terms of training efficiency and accuracy. These improvements highlight the suitability of the proposed framework for real-time control scenarios, where computational efficiency and rapid adaptation are critical.

The integration of interactive web-based learning media is essential to address challenges in teaching hydrocarbon materials at SMAN 1 Sungai Pinang. Traditional teaching methods, while effective, often lack student engagement, motivation, and interactivity. This study explores the need for digital learning tools to enhance chemistry education. A needs analysis was conducted through surveys and questionnaires to assess student and teacher perceptions regarding existing teaching methods. The validity and reliability of the questionnaire were confirmed to ensure accurate data collection. The study also examined school facilities and academic challenges that impact the implementation of interactive learning. Findings indicate that well-equipped school facilities support the adoption of interactive and digital learning strategies. While traditional methods remain beneficial, a blended learning approach is necessary to optimize student engagement and academic performance. The study also highlights various academic and non-academic challenges faced by students, necessitating further support and development. Regular evaluation of teaching strategies is crucial to adapting to students' evolving needs. The study underscores the importance of interactive web-based learning media in improving teaching effectiveness and student engagement. The findings suggest that supplementing traditional methods with digital tools can enhance comprehension and motivation. This study supports the development of interactive digital learning media to complement traditional chemistry instruction. Implementing blended learning strategies can address current educational challenges, fostering a more engaging and effective learning environment.

Traffic signal control is a complex problem and it is difficult to determine an optimal strategy to control multi-directional traffic at multiple intersections. Recent years have witnessed numerous successes of deep learning neural networks in the fields of artificial intelligence. Motivated by the dominant performance of neural networks, this study attempts to develop a novel adaptive signal control approach by fusing deep learning (DL) and reinforcement learning (RL), i.e., deep reinforcement learning (DRL), for arterial signal coordination. DRL can considerably improve the ability to deal with large amounts of data processing, systematic perception and expression, which is key to coordinated control of arterial intersections. The proposed algorithm is implemented by utilizing real-time traffic detection data and aims to optimize the hybrid global and local reward functions. The experimental results obtained by traffic simulation software SUMO demonstrate the advantage of the proposed approach, as well as its efficiency and effectiveness compared with fixed-time and actual signal control methods.

Teaching descriptive texts at the junior high school level, particularly in Grade 7, often encounters challenges in creating an interactive and inclusive learning environment, especially for students with special needs. This study aims to explore and analyze effective interactive learning strategies for teaching descriptive texts, with a particular focus on technology integration, adaptations for special needs students, and comprehensive evaluation methods. Using a literature review approach, this article examines recent sources from educational journals, textbooks, and scholarly publications in the fields of Indonesian language instruction, educational technology, and inclusive education. The discussion addresses fundamental concepts of descriptive texts, interactive learning strategies such as the use of visual and audio media, group activities, language games, and the integration of technology into the learning process. The findings indicate that interactive learning approaches which incorporate technology and address students’ diverse needs significantly enhance descriptive writing skills, creativity, and student engagement. In conclusion, implementing interactive learning strategies in teaching descriptive texts not only improves students' academic abilities but also fosters critical thinking and social skills, ultimately creating a more inclusive and meaningful learning experience for all students.

Interactive Machine Learning (IML) can enable intelligent systems to interactively learn from their end-users, and is quickly becoming more and more relevant to many application domains. Although it places the human in the loop, interactions are mostly performed via mutual explanations that miss contextual information. Furthermore, current model-agnostic IML strategies such as CAIPI are limited to ’destructive’ feedback, meaning that they solely allow an expert to prevent a learner from using irrelevant features. In this work, we propose a novel interaction framework called Semantic Interactive Learning for the domain of document classification, located at the intersection between Natural Language Processing (NLP) and Machine Learning (ML). We frame the problem of incorporating constructive and contextual feedback into the learner as a task involving finding an architecture that enables more semantic alignment between humans and machines while at the same time helping to maintain the statistical characteristics of the input domain when generating user-defined counterexamples based on meaningful corrections. Therefore, we introduce a technique called SemanticPush that is effective for translating conceptual corrections of humans to non-extrapolating training examples such that the learner’s reasoning is pushed towards the desired behavior. Through several experiments we show how our method compares to CAIPI, a state of the art IML strategy, in terms of Predictive Performance and Local Explanation Quality in downstream multi-class classification tasks. Especially in the early stages of interactions, our proposed method clearly outperforms CAIPI while allowing for contextual interpretation and intervention. Overall, SemanticPush stands out with regard to data efficiency, as it requires fewer queries from the pool dataset to achieve high accuracy.

We propose Dynamic-goal Deep Reinforcement Learning (DGDRL) method to address the problem of robot arm motion planning in telemanipulation applications. This method intuitively maps human hand motions to a robot arm in real-time, while avoiding collisions, joint limits and singularities. We further propose a novel hardware setup, based on the HTC VIVE VR system, that enables users to smoothly control the robot tool position and orientation with hand motions, while monitoring its movements in a 3D virtual reality environment. A VIVE controller captures 6D hand movements and gives them as reference trajectories to a deep neural policy network for controlling the robot’s joint movements. Our DGDRL method leverages the state-of-art Proximal Policy Optimization (PPO) algorithm for deep reinforcement learning to train the policy network with the robot joint values and reference trajectory observed at each iteration. Since training the network on a real robot is time-consuming and unsafe, we developed a simulation environment called RobotPath which provides kinematic modeling, collision analysis and a 3D VR graphical simulation of industrial robots. The deep neural network trained using RobotPath is then deployed on a physical robot (ABB IRB 120) to evaluate its performance. We show that the policies trained in the simulation environment can be successfully used for trajectory planning on a real robot. The the codes, data and video presenting our experiments are available at https://github.com/kavehkamali/ppoRobotPath.

Interactive machine learning systems allow end-users, often non-experts, to build and apply statistical models for their own uses. Constructivism is the view that learning occurs when ideas and experiences interact. I argue that the objectives of interactive machine learning can be interpreted as constructivist. By so characterising them, I show how constructivist learning environments pose critical questions for the design of interactive machine learning systems.

The object of the study is a mobile walking robot with two or more movable articulated limbs. The concept of a "critical condition" is introduced, in which the mechanism balances on the verge of falling (but does not fall) or there is a possibility of damage to mechanical components due to the generation of unacceptable joint angles. The subject of the study is a general algorithm for the elimination of critical conditions, which provides the possibility of training an agent based on a deep reinforcement learning algorithm directly on a real robot, without the risk of damaging its mechanisms and interrupting the process of interaction with the environment to restore a stable state. The purpose of this work is to develop a general algorithm for the elimination of critical conditions in the context of adaptive control of a walking robot based on deep learning algorithms with reinforcement. A comparison was made between the proposed and standard methods of applying deep OP on a real robot. The experiments were conducted on 6,000 episodes, with a dimension of 300 steps each. The following quality metrics were selected for evaluation: the percentage of episodes without an actual fall, the percentage of fully completed episodes, and the maximum episode length. The algorithm is based on the concept of "critical condition" and uses the following principles and methods: the "trial and error" method, the feedback principle, holding the projection of the center of gravity point in the area of the polygon formed by the points of contact of the limbs with the work surface, which ensures the balancing of the structure and allows you to determine the boundary areas in which the robot is still stable. The scientific novelty of the work lies in the proposed approach, which allows an intelligent agent to control a physical robot "directly", without pre-configuration in a simulation environment with subsequent transfer implementation. The proposed algorithm is not aimed at improving the agent's performance, but is intended to provide greater autonomy in the learning process of the robot, directly in the hardware. The basic idea is to immediately respond to a critical condition in the form of the fastest sequential return to a certain number of steps back along the decision-making trajectory, ensuring that the agent remains in a stable and safe state at all times. The method of proximal policy optimization (PPO) was used as a method of deep reinforcement learning. As a result of the comparative analysis, the proposed algorithm demonstrated a hundredfold increase in the stability of the mechanism.