In this research, we propose a new self-sufficient bicycle sharing operating system by local residents. The system incorporates Deep Q-Networks (DQN) at each station and uses a Multi-Agent Reinforcement Learning (MARL) model. In the proposed model, multiple reinforcement learning agents collaborate and provide incentives to local residents which can avoid making bicycles redistributed unevenly to bicycle stations. Additionally, we compare the simple method using the number of remaining bicycles, the MARL model, and a single-agent reinforcement learning (SARL) model to verify the improvement in learning speed and flexibility to changes in the environment. The simulation results show that the MARL model reduces the learning time due to the collaborative actions of multiple agents and results in more efficient service operations compared to the simple method and the SARL model.

Ginger rhizome rot is one of the most damaging plant diseases that occur in ginger. In this study, we proposed a method of detecting diseased ginger plants by using leaf motion to identify diseased ginger plants at an early stage of rhizome rot. We found a tendency for inoculated plants intentionally inoculated with the rhizome rot pathogen to have less leaf motion than uninoculated plants. Hereby, we proposed a method based on ExG (Excess Green) and SVM (Support Vector Machine) to quantify leaf motion and evaluated whether these methods could discriminate between inoculated and uninoculated plants using ROC (Receiver Operator Characteristic) curves and AUC (Area Under the Curve). As a result, the AUC was 0.989, and the accuracy for discrimination by the best threshold in Youden's Index was 96.08%. The true positive rate was 92.59%. The false positive rate was 0% when the inoculated plants were considered positive. When the threshold value was set so that the true positive rate was 100%, the minimum false positive rate was 16.67%.

Driving in roads shared with pedestrians remains to be a challenging research problem in suburban/urban autonomous driving scenario. Interacting with pedestrians and making considerate decisions require high level decision making which is close to that of humans. This paper attempts to propose a multi-mode pedestrian model and a model predictive control (MPC) framework that uses the pedestrian model to predict the interactive behavior of the pedestrian. A probability weighted ARX (PrARX) model is used as the pedestrian model. This model can represent both decision making and motion of the pedestrian considering the interaction with ego car. The allowed risk levels of the controller is tunable to create a more personal driving behavior. The decision entropy of the pedestrian is also included as a cost factor in the MPC. This makes the pedestrian's decision making process easier and leads to realization of considerate driving. The validation of proposed model is performed based on the vehicle-crowd interaction (VCI) - CITR dataset and the performance of the controller is demonstrated in simulations using MATLAB.