|
Multi-agent Reinforcement Learning in a Large Scale Environment via Supervisory Network and Curriculum Learning |
|
Seungwon Do(ETRI, Korea), Changeun Lee(Electronics and Telecommunications Research Institute (ETRI), Korea) |
| Multi-agent reinforcement learning is essential for optimizing policy for collaboration and competition environments. However, as the action space of the agent increases, the number of state-action pairs which have to be explored increases exponentially. To solve this problem, we propose a supervisory network. To achieve the global goal, the supervisory network creates a sub-goal and assigns the goals to the agents so that the agents can effectively learn the optimal policy with a small action space. In addition, we adapt the curriculum learning method to learn a large-scale environment. As a consequence, the agent can explore the environment in which the complexity increases gradually. |
|
|
Neuro-evolutionary based Controller Design for Linear and Non-linear Systems |
|
Samarth Singh, Kaushal Kishore, S A Akbar(CSIR-CEERI, India) |
| In the present work a Neuro-Evolution based approach has been used to train a neural network for control of some sample systems. This method makes use of Genetic algorithm, here it is generating a population of neural networks and introduces mutation for producing better off-springs for the next generation. It makes use of fitness function to evaluate performance of off-springs, this fitness function is based on a novel reward function which allows for quick and smooth settling of the sample system towards set point. In order to address dynamics of the system's time sequenced error has been been taken as exogenous input for the neural network. Method tested on linear and non linear systems. |
|
|
Intra-Batch Features Separation for Indoor and Outdoor Pedestrian-View Intersection Classification |
|
Marcella Astrid, Muhammad Zaigham Zaheer(University of Science and Technology, Korea), Seung-Ik Lee(Electronics and Telecommunications Research Institute, Korea) |
| To solve pedestrian-view intersection classification problem, previous approaches simply fine-tune an ImageNet-pretrained network with intersection classification dataset using cross-entropy loss. In this work, we propose a novel additional loss to further improve the model's capability to discriminate intersection and non-intersection class. This loss is directly calculated on the features in a given mini-batch without requiring any additional inference. Furthermore, previous works cover only outdoor domain while we also propose indoor domain in addition to the outdoor intersection classification dataset. |
|
|
Weight Change Prediction for Automated Depression Diagnosis |
|
Juyoung Hong, Jeongmin Shin, Yujin Hwang, Jeongmin Lee, Yukyung Choi(Sejong University, Korea) |
| We have presented a novel method that classified weight changes using a facial image for self-diagnosis of depression. This method inferred weight change by utilizing geometry information from face mesh. We have also introduced powerful pre-processing, which are 3D face alignment and monocular depth estimation to make the model predict weight change successfully in various viewpoints and distances. |
|
|
Physical Violence Detection in Video Streaming Using Partitioned Skeleton Analysis |
|
Batyrkhan Omarov(Al-Farabi Kazakh National University, Kazakhstan), Sergazy Narynov, Zhandos Zhumanov, Aidana Gumar, Mariyam Khassanova(Alem Research, Kazakhstan) |
| We propose a skeleton-based method for identifying hostile behavior in this paper. The method does not require a lot of powerful hardware, but it is very quick to implement. Our approach consists of two stages: feature extraction from video frames to assess a person's posture, followed by action classification using a neural network to identify whether the frames include bullying situations. We also generated a dataset of 400 minutes of video data comprising one person's activities and 20 hours of video data including physical bullying and aggressive acts, as well as 13 classifications for distinguishing aggressor and victim behavior. On the gathered dataset, the method was put to the test. |
|
|
Framework for Deep Learning-enabled Building High-Definition Maps and Planning Routes for Autonomous Vehicles |
|
Daegyu Lee, Hyungi Kim, D.Hyunchul Shim(KAIST, Korea) |
| In this study, we proposed a deep learning-enabled HD map-building framework and route planning algorithm.
To build an accurate HD map, we proposed LiDAR scan matching algorithm for localization; Furthermore, based on this map, a route planning algorithm was implemented to enable the vehicle to arrive at its destination via the optimal route. We decreased the positional estimation error by 26.5% and improved the scan matching algorithm speed by 17.4% compared to using a conventional filtering algorithm. We automatically generated a two-dimensional (2-D) HD map using deep learning with an average distance error of 0.3639 m over a travel distance of 5,880 m. |
|