The ability to make informed decisions in complex scenarios is crucial for intelligent automotive systems. Traditional expert rules and other methods often fall short in complex contexts. Recently, reinforcement learning has garnered significant attention due to its superior decision-making capabilities. However, there exists the phenomenon of inaccurate target network estimation, which limits its decision-making ability in complex scenarios. This paper mainly focuses on the study of the underestimation phenomenon, and proposes an end-to-end autonomous driving decision-making method based on an improved TD3 algorithm. This method employs a forward camera to capture data. By introducing a new critic network to form a triple-critic structure and combining it with the target maximization operation, the underestimation problem in the TD3 algorithm is solved. Subsequently, the multi-timestep averaging method is used to address the policy instability caused by the new single critic. In addition, this paper uses Carla platform to construct multi-vehicle unprotected left turn and congested lane-center driving scenarios and verifies the algorithm. The results demonstrate that our method surpasses baseline DDPG and TD3 algorithms in aspects such as convergence speed, estimation accuracy, and policy stability.

Accurate prediction of passenger demand is essential for the efficient operation and management of metro systems. In practical scenarios, strategies to enhance metro service quality often require passenger demand information on multiple fronts, such as inflow to a station, outflow from a station, as well as transition flow between entry/exit stations. While predictions for a single type of passenger demand have been extensively studied, limited attention was paid to jointly predicting multiple demands. This problem is challenging due to the complex relationships among multiple demands (e.g., inflow is only correlated with historical inflow, while the outflow is not only correlated with outflow but also determined by the inflow) and the imbalanced training issue of multiple prediction tasks. To address these challenges, this paper proposes a deep multi-task learning (MTL) model called MTLMetro to co-predict multiple demands in metro systems. More specifically, we deploy the message-passing schemes in graph neural networks (GNNs) as the knowledge-sharing mechanisms in the MTL model to capture the inherent relationships among multiple demands. To balance the training of multiple tasks, we introduce a novel weighting scheme named dynamic weight average (DWA), which can dynamically adapt relative weight for each task. In addition, the partial observability problem of transition flow is also considered in MTLMetro in an end-to-end manner. Empirical evaluation on a real-world dataset demonstrates MTLMetro’s superior performance across the different demand prediction tasks when compared to several benchmarks. Further ablation experiments verify the effectiveness of the proposed modules and the weighting method.

With the increasing importance of subways in urban public transportation systems, pedestrian flow simulation for supporting station management and risk analysis becomes more necessary. There is a need to calibrate the simulation model parameters with real-world pedestrian flow data to achieve a simulation closer to the real situation. This study presents a calibration approach based on YOLOv5 for calibrating the simulation model parameters in the social force model inserted in Anylogic. This study compared the simulation results after model calibration with real data. The results show that (1) the parameters calibrated in this paper can reproduce the characteristics of pedestrian flow in the station; (2) the calibration model not only decreases global errors but also overcomes the common phenomenon of large differences between simulation and reality.

This paper proposes an integrated optimization model for vessel scheduling and routing. The objective is to maximize shipping company profits while considering profit volatility using the Conditional Value-at-Risk metric to master risks from demand fluctuations. Simultaneously, the model balances the spot and contract container allocation by optimally adjusting shipping speeds so as to minimize carbon emissions. We account for vessel deployment, chartering costs, delay penalties, fuel expenses, and weather conditions to ensure the model’s compatibility with the practical transporting environment. In particular, a hybrid demand prediction model, combining long short-term memory and multi-scale network techniques, predicts spot and contract container volumes at ports, facilitating real-time allocation and more precise scheduling optimization. Two hybrid heuristics, one adaptive large-neighborhood search algorithm, and the Gurobi solver are devised and compared based on the efficiency and accuracy of solving the model. The results indicate that our optimization offers practical insights for shipping companies, enabling them to achieve a better trade-off between profits and risks, promoting a promising maritime transport career.