|
Bounds on the Number of Training Samples Needed for Inverse Kinematics Solutions by Artificial Neural Networks |
|
Dongwon Lim(University of Suwon, Korea) |
| The inverse kinematics (IK) practice that only a few points (training samples) of the end effector are recorded can be planned to operate a manipulator by the Artificial Neural Networks (ANN). More training samples and hidden neurons are, better the IK function fitted by ANN performs. However, with a high number of samples and hidden neurons, the procedure becomes impractical. This paper attempts to provide a mathematical framework for a reasonable number of ANN training samples for acceptable operations. Mathematical bound estimates are derived between trained and untrained points. Through simulation studies, errors by bound estimates and by ANN were compared. |
|
|
Model-Based Reinforcement Learning for Environments with Delayed Feedback |
|
Jangwon Kim, Hangyeol Kim, Jiwook Kang, Soohee Han(POSTECH, Korea) |
| Reinforcement learning (RL) has recently advanced in games, control tasks, and real-world applications. However, applying RL to reality is often challenging due to inevitable signal delays. These lead to performance degradation due to the discrepancy between delayed observation and true observation. To address this, we introduce the Model-Based State Estimation (MBSE) algorithm. In MBSE, the agent uses a learned transition model to estimate the true current state, guiding optimal action selection. We tested our algorithm on MuJoCo control tasks and compared it with other algorithms in a delayed feedback environment, and our algorithm showed significant performance improvement. |
|
|
Deep Neural Network Approach for Automated Architectural Bolt Usage Prediction in Building Information Model Control |
|
Soohee Han, Jonghyeok Park(POSTECH, Korea) |
| This paper introduces a data-driven DNN model designed to automate Building Information Model (BIM) control by accurately predicting architectural bolt usage. The model was trained using a substantial dataset consisting of 13,000 samples. The validation results achieved high performance, with an average accuracy surpassing 90% for both the x-axis and y-axis data. These achieved accuracy levels are notably high, signifying the model's suitability for real-world BIM controllers. |
|
|
Recovery Direction Classification: Reduced Order Quadrotor Dynamics Based Method |
|
Soohee Han, Changhyeon Lee, Woojoung Lim(POSTECH, Korea), Seunghyeon Sim, Jooyeol Jung(Convergence IT Engineering, Korea) |
| This paper proposes a reduced dynamics-based recovery direction classification method. By simplifying quadrotor dynamics into a 1 dimensional representation and predicting overall recovery time using the Newton-Raphson method, we enable swift operation in critical recovery scenarios, enhancing quadrotor recovery control. |
|
|
Augmented-Ensemble TD3 : Overcoming the Shackles of Constant Action Delay |
|
Soohee Han, Jongsoo Lee(POSTECH, Korea) |
| Reinforcement Learning has experienced significant advances in various domains. However, delayed feed-
back in RL environments poses challenges due to the violation of the Markovian property. In this paper, we propose
an approach to address the issues of Markov Decision Process(MDP) with delayed feedback. The proposed approach,
called ”Augmented-Ensemble Twin-Delayed Deep Deterministic Policy Gradient(TD3),” aims to mitigate the perfor-
mance degradation caused by delayed feedback. |
|
|
Unveiling the Potential and Limitations of Visual Reinforcement Learning: An Experimental Analysis on the Cart-Pole Inverted Pendulum System |
|
Soohee Han, Sanghyun Ryoo(POSTECH, Korea) |
| Reinforcement learning can handle what traditional control theory can not cope with. In the field of visual reinforcement learning given only a partially observed high-dimensional image state, the agent still can learn a policy to achieve certain tasks. In this paper, we show the potential and limitations of visual reinforcement learning with an experiment on cart-pole balancing tasks both in simulation and semi-real environment systems. |
|