In real-world recommender model deployments, the models are typically retrained and deployed repeatedly. It is the rule-of-thumb to periodically retrain recommender models to capture up-to-date user behavior and item trends. However, the harm caused by delayed model updates has not been investigated extensively yet. in this perspective paper, we formulate the delayed model update problem and quantitatively demonstrate the delayed model update actually harms the model performance by increasing the number of cold users and cold items increase and decreasing overall model performances. These effects vary across different domains having different characteristics. Upon these findings, we further argue that although the delayed model update has negative effects on online recommender model deployment, yet it has not gathered enough attention from research communities. We argue our verification of the relationship between the model update cycle and model performance calls for further research such as faster model training, and more efficient data pipelines to keep the model more up-to-date with the latest user behaviors and item trends.

Recently, deep reinforcement learning (RL) has proven its feasibility in solving combinatorial optimization problems (COPs). The learning-to-rank techniques have been studied in the field of information retrieval. While several COPs can be formulated as the prioritization of input items, as is common in the information retrieval, it has not been fully explored how the learning-to-rank techniques can be incorporated into deep RL for COPs. In this paper, we present the learning-to-rank distillation-based COP framework, where a high-performance ranking policy obtained by RL for a COP can be distilled into a non-iterative, simple model, thereby achieving a low-latency COP solver. Specifically, we employ the approximated ranking distillation to render a score-based ranking model learnable via gradient descent. Furthermore, we use the efficient sequence sampling to improve the inference performance with a limited delay. With the framework, we demonstrate that a distilled model not only achieves comparable performance to its respective, high-performance RL, but also provides several times faster inferences. We evaluate the framework with several COPs such as priority-based task scheduling and multidimensional knapsack, demonstrating the benefits of the framework in terms of inference latency and performance.

In this paper, we propose a Geometric Model-Free Approach with a Particle Filter (GMFA-PF) through the use of automotive LiDAR for real-time tracking of moving objects within an urban driving environment. GMFA-PF proved to be lightweight, capable of finishing the process within the sensing period of the LiDAR on a single CPU. The proposed GMFA-PF tracks and estimates moving objects without any assumptions on the geometry of the target. This approach enables efficient tracking of multiple object classes, with robustness to a sparse point cloud. Point cloud on moving objects is classified via the predicted Static Obstacle Map (STOM). A likelihood field is generated through the classified point cloud and is used in particle filtering to estimate the moving object’s pose, shape, and speed. Quantitative and qualitative comparisons - with Geometric Model-Based Tracking (MBT), Deep Neural Network (DNN), and GMFA - are performed for GMFA-PF using urban driving and scenario driving data gathered on an autonomous vehicle fitted with close-to-market sensors. The proposed approach shows robust tracking and accurate estimation performance in both sparse and dense point clouds; GMFA-PF achieves improved tracking performance in dense traffic and reduces yaw estimation delay compared to the others. Autonomous vehicles with GMFA-PF demonstrated auto-nomous driving on urban roads.

This paper proposes a lane marking detection method for automated driving in complex urban roads to be used with map-matching localization algorithms. First, a LiDAR based lane detection and map matching algorithm is explained and a lane marking detection algorithm using AVM (Around View Monitor) cameras is presented. The AVM camera based lane marking detection algorithm includes camera calibration, bird-eye view conversion, and lane marking detection. The two lane detection methods are combined and implemented as a part of a map-matching localization algorithm on an autonomous test vehicle. Multiple experiments on various test routes, including complex situations like driver's license exam stations, were performed to verify the performance of the lane marking detection in the map-matching algorithm.

Controller Area Network (CAN) is a bus communication protocol which defines a standard for reliable and efficient transmission between in-vehicle nodes in real-time. Since CAN message is broadcast from a transmitter to the other nodes on a bus, it does not contain information about the source and destination address for validation. Therefore, an attacker can easily inject any message to lead system malfunctions. In this paper, we propose an intrusion detection method based on the analysis of the offset ratio and time interval between request and response messages in CAN. If a remote frame having a particular identifier is transmitted, a receiver node should respond to the remote frame immediately. In attack-free state, each node has a fixed response offset ratio and time interval while these values vary in attack state. Using this property, we can measure the response performance of the existing nodes based on the offset ratio and time interval between request and response messages. As a result, our methodology can detect intrusions by monitoring offset ratio and time interval, and it allows quick intrusion detection with high accuracy.

In this paper, a hidden Markov model (HMM)-based dynamic sensor scheduling problem is formulated, and solved using rollout concepts to overcome the computational intractability of the dynamic programming (DP) recursion. The problem considered here involves dynamically sequencing a set of sensors to minimize the sum of sensor cost and the HMM state estimation error cost. The surveillance task is modeled as a single HMM with multiple emission matrices corresponding to each of the sensors. The rollout information gain (RIG) algorithm proposed herein employs the information gain (IG) heuristic as the base algorithm. The RIG algorithm is illustrated on an intelligence, surveillance, and reconnaissance (ISR) scenario of a village for the presence of weapons and terrorists/refugees. Extension of the RIG strategy to monitor multiple HMMs involves combining the information gain heuristic with the auction algorithm that computes the κ-best assignments at each decision epoch of rollout.