The job scheduler plays a vital role in high-performance computing platforms. It determines the execution order of the jobs and the allocation of resources, which in turn affect the resource utilization of the entire system. As the scale and complexity of HPC continue to grow, job scheduling is becoming increasingly important and difficult. Existing studies relied on user-specified or regression techniques to give fixed runtime prediction values and used the values in static heuristic scheduling algorithms. However, these approaches require very accurate runtime predictions to produce better results, and fixed heuristic scheduling strategies cannot adapt to changes in the workload. In this work, we propose RLSchert, a job scheduler based on deep reinforcement learning and remaining runtime prediction. Firstly, RLSchert estimates the state of the system by using a dynamic job remaining runtime predictor, thereby providing an accurate spatiotemporal view of the cluster status. Secondly, RLSchert learns the optimal policy to select or kill jobs according to the status through imitation learning and the proximal policy optimization algorithm. Extensive experiments on real-world job logs at the USTC Supercomputing Center showed that RLSchert is superior to static heuristic policies and outperforms the learning-based scheduler DeepRM. In addition, the dynamic predictor gives a more accurate remaining runtime prediction result, which is essential for most learning-based schedulers.

Deep Reinforcement Learning (DRL) is a promising approach for general artificial intelligence. However, most DRL methods suffer from the problem of data inefficiency. To alleviate this problem, DeepMind proposed Prioritized Experience Replay (PER). Though PER improves data utilization, the priorities of most samples in its Experience Memory (EM) are out of date, as only the priorities of a small part of the data are updated while the Q network parameters are updated. Consequently, the difference between storage and real priority distributions gradually increases, which will introduce bias into the gradients of Deep Q-Learning (DQL) and make the DQL update toward a non-ideal direction. In this work, we propose a novel self-adaptive priority correction algorithm named Importance-PER (Imp-PER) to fix the update deviation. Specifically, we predict the sum of real Temporal-Difference error (TD-error) of all data in EM. Data are corrected by an importance weight, which is estimated by the predicted sum and the real TD-error calculated by the latest agent. To control the unbounded importance weight, we use truncated importance sampling with a self-adaptive truncation threshold. The conducted experiments on various games of Atari 2600 with Double Deep Q-Network and MuJoCo with Deep Deterministic Policy Gradient demonstrate that Imp-PER improves the data utilization and final policy quality on discrete states and continuous states tasks without increasing the computational cost.

Spatial reasoning in Large Language Models (LLMs) serves as a foundation for embodied intelligence. However, even in simple maze environments, LLMs often struggle to plan correct paths due to hallucination issues. To address this, we propose S2ERS, an LLM-based technique that integrates entity and relation extraction with the on-policy reinforcement learning algorithm Sarsa for optimal path planning. We introduce three key improvements: (1) To tackle the hallucination of spatial, we extract a graph structure of entities and relations from the text-based maze description, aiding LLMs in accurately comprehending spatial relationships. (2) To prevent LLMs from getting trapped in dead ends due to context inconsistency hallucination by long-term reasoning, we insert the state-action value function Q into the prompts, guiding the LLM’s path planning. (3) To reduce the token consumption of LLMs, we utilize multi-step reasoning, dynamically inserting local Q-tables into the prompt to assist the LLM in outputting multiple steps of actions at once. Our comprehensive experimental evaluation, conducted using closed-source LLMs ChatGPT 3.5, ERNIE-Bot 4.0 and open-source LLM ChatGLM-6B, demonstrates that S2ERS significantly mitigates the spatial hallucination issues in LLMs, and improves the success rate and optimal rate by approximately 29% and 19%, respectively, in comparison to the SOTA CoT methods.

In the context where social media emerges as a pivotal platform for social movements and shaping public opinion, accurately simulating and predicting the dynamics of user opinions is of significant importance. Such insights are vital for understanding social phenomena, informing policy decisions, and guiding public opinion. Unfortunately, traditional algorithms based on idealized models and disregarding social data often fail to capture the complexity and nuance of real-world social interactions. This study proposes the Fusing Dynamics Equation-Large Language Model (FDE-LLM) algorithm. This innovative approach aligns the actions and evolution of opinions in Large Language Models (LLMs) with the real-world data on social networks. The FDE-LLM divides users into two roles: opinion leaders and followers. Opinion leaders use LLM for role-playing and employ Cellular Automata(CA) to constrain opinion changes. In contrast, opinion followers are integrated into a dynamic system that combines the CA model with the Susceptible-Infectious-Recovered (SIR) model. This innovative design significantly improves the accuracy of the simulation. Our experiments utilized four real-world datasets from Weibo. The result demonstrates that the FDE-LLM significantly outperforms traditional Agent-Based Modeling (ABM) algorithms and LLM-based algorithms. Additionally, our algorithm accurately simulates the decay and recovery of opinions over time, underscoring LLMs potential to revolutionize the understanding of social media dynamics.

Resource demands prediction has become a promising tool to facilitate automatic scaling of resource management, which makes it feasible to reduce the cost and improve resource utilization in the cloud. Most current prediction methods of resource demands are based on a single model. However, the resource load patterns are usually diversity and variability in the cloud, it is hard to get the satisfactory prediction performance through a single model. To solve this problem, LSRP is proposed in this paper which is a novel ensemble approach for resource demands prediction. The method contains two important learning phases. First, we use a set of sub-models to predict the resource demands respectively based on the historic workloads. Second, we regard the results of sub-models as inputs to a second-level model. This second-level model is a linear model which is trained to combine the sub-models. Extensive experiments on production data set demonstrate that the proposed model outperforms the sub-models and other ensemble models in accuracy and adaptation.