Retrieval-Augmented Generation (RAG) critically depends on effective query expansion to retrieve relevant information. However, existing expansion methods adopt uniform strategies that overlook user-specific semantics, ignoring individual expression styles, preferences, and historical context. In practice, identical queries in text can express vastly different intentions across users. This representational rigidity limits the ability of current RAG systems to generalize effectively in personalized settings. Specifically, we identify two core challenges for personalization: 1) user expression styles are inherently diverse, making it difficult for standard expansions to preserve personalized intent. 2) user corpora induce heterogeneous semantic structures—varying in topical focus and lexical organization—which hinders the effective anchoring of expanded queries within the user’s corpora space. To address these challenges, we propose Personalize Before Retrieve (PBR), a framework that incorporates user-specific signals into query expansion prior to retrieval. PBR consists of two components: P-PRF, which generates stylistically aligned pseudo feedback using user history for simulating user expression style, and P-Anchor, which performs graph-based structure alignment over user corpora to capture its structure. Together, they produce personalized query representations tailored for retrieval. Experiments on two personalized benchmarks show that PBR consistently outperforms strong baselines, with up to 10% gains on PersonaBench across retrievers. Our findings demonstrate the value of modeling personalization before retrieval to close the semantic gap in user-adaptive RAG systems.

Listwise reranking with Large Language Models (LLMs) has emerged as the state-of-the-art approach, consistently establishing new performance benchmarks in passage reranking. However, their practical application faces two critical hurdles: the prohibitive computational overhead and high latency of processing long token sequences, and the performance degradation caused by phenomena like "lost in the middle" in long contexts. To address these challenges, we introduce Compress-then-Rank (C2R), an efficient framework that performs listwise reranking not on original passages, but on their compact multi-vector surrogates. These surrogates can be pre-computed and cached for all passages in the corpus. The effectiveness of C2R hinges on three key innovations. First, the compressor model is pre-trained on a combination of text restoration and continuation objectives, enabling high-fidelity compressed vector sequences that mitigate the semantic loss common in single-vector methods. Second, a novel input scheme prepends embeddings of each ordinal index (e.g., [1]:) to its corresponding compressed vector sequence, which both delineates passage boundaries and guides the reranker LLM to generate a ranked list. Finally, the compressor and reranker are jointly optimized, making the compression explicitly ranking-aware for the ranking objective. Extensive experiments on major reranking benchmarks demonstrate that C2R provides substantial speedups while achieving competitive and even superior ranking performance compared to full-text reranking methods.

Cloud-device collaboration leverages on-cloud Large Language Models (LLMs) for handling public user queries and on-device Small Language Models (SLMs) for processing private user data, collectively forming a powerful and privacy-preserving solution.However, existing approaches often fail to fully leverage the scalable problem-solving capabilities of on-cloud LLMs while underutilizing the advantage of on-device SLMs in accessing and processing personalized data.This leads to two interconnected issues: 1) Limited utilization of the problem-solving capabilities of on-cloud LLMs, which fail to align with personalized user-task needs, and 2) Inadequate integration of user data into on-device SLM responses, resulting in mismatches in contextual user information.In this paper, we propose a Leader-Subordinate Retrieval framework for Privacy-preserving cloud-device collaboration (LSRP), a novel solution that bridges these gaps by: 1) enhancing on-cloud * Contributed equally to this work.

Path planning is an essential task for the mission execution of unmanned surface vehicle (USV). However, existing advanced techniques based on deep reinforcement learning (DRL) often suffer from low learning efficiency and insufficient environmental perception from single-sensor configurations. To address these issues, this paper proposes a hybrid framework named A*D3QN, which integrates the heuristic efficiency of the A* algorithm with the adaptive decision-making of a dueling double deep Q-network (D3QN), associated with the multimodal data fusion for precise environment modeling. The proposed A*D3QN incorporates prior knowledge from the global paths generated by A* to initialize and guide the D3QN learning process. The prior knowledge, formatted as RL transition tuples, is used in the reward function design and the N-step prioritized experience replay, which significantly accelerates overall learning efficiency. Moreover, an improved D3QN architecture is designed to dynamically fuse visual data and navigation states via a cross-entropy attention mechanism, enabling multimodal perception in partially unknown environments. Extensive experiments across three scenarios with varying obstacle densities demonstrate that A*D3QN significantly outperforms state-of-the-art DRL baselines. Ablation studies further validate the necessity of each component.

House layout plays a crucial role in housing planning and design. In recent years, automated generation of house layouts has gained significant attention. The objective is to automatically generate floorplans that meet specific requirements under given constraints. In this paper, we present an extension and improvement of the existing diffusion model, focusing on the generation of housing layouts in more complex constrained scenarios. Firstly, we introduce the consideration of relative room area ranking as a new problem scenario. Secondly, we propose an indirect encoding approach that represents the ranking relationship of room relative areas as a directed graph to address potential issues arising from embedding ranking features. Finally, we introduce a corresponding attention module to capture the newly added constraint relationships. Moreover, we evaluate the proposed approach using a range of metrics and the RPLAN dataset. Our proposed method demonstrates improved accuracy in considering variations in room sizes and provides more reasonable layout. The obtained results also indicate that our enhanced approach exhibits better compatibility and Spearman correlation for relative room area ranking compared to the state-of-the-art methods.

The ability to swiftly and precisely reach any user-specified target location is necessary for a robotic arm that can be used in real-world scenarios. To date, many evolutionary optimization algorithms have been used to design controllers for robotic arms. However, when designing a robotic arm to reach multiple targets, most existing methods need to evolve the control strategy from scratch for each target, rather than trying to reuse existing experience. Therefore, computational resources are repeatedly and meaninglessly consumed. To this end, this paper proposes a genetic algorithm based on knowledge transfer (GAKT) dedicated to reusing existing knowledge to optimize a new robotic arm control task. Specifically, the knowledge transfer process can be summarized into the following two steps. First, through sequential transfer, GAKT initializes the population with the help of a knowledge base constructed by a quality diversity algorithm. Second, underperforming individuals are encouraged to acquire knowledge from excellent individuals in the same generation during the optimization process. We tested the effectiveness of GAKT and investigated its average performance by selecting multiple target points in different dimensions. The results show that GAKT can find the most advantageous arrival strategy (that is, make the end of the manipulator the closest to the target) on most of the selected targets. Moreover, we conducted ablation experiments and demonstrated the effectiveness of the knowledge transfer processes.

Users in recommender systems exhibit multi-behavior in multiple business scenarios on real-world e-commerce platforms. A crucial challenge in such systems is to make recommendations for each business scenario at the same time. On top of this, multiple predictions (e.g., Click Through Rate and Conversion Rate) need to be made simultaneously in order to improve the platform revenue. Research focus on making recommendations for several business scenarios is in the field of Multi-Scenario Recommendation (MSR), and Multi-Task Recommendation (MTR) mainly attempts to solve the possible problems in collaboratively executing different recommendation tasks. However, existing researchers have paid attention to either MSR or MTR, ignoring the integration of MSR and MTR that faces the issue of conflict between scenarios and tasks. To address the above issue, we propose a Meta-based Multi-scenario Multi-task RECommendation framework (M3REC) to serve multiple tasks in multiple business scenarios by a unified model. However, integrating MSR and MTR in a proper manner is non-trivial due to: 1) Unified representation problem: Users’ and items’ representation behave Non-i.i.d in different scenarios and tasks which takes inconsistency into recommendations. 2) Synchronous optimization problem: Tasks distribution varies in different scenarios, and a unified optimization method is needed to optimize multi-tasks in multi-scenarios. Thus, to unified represent users and items, we design a Meta-Item-Embedding Generator (MIEG) and a User-Preference Transformer (UPT). The MIEG module can generate initialized item embedding using item features through meta-learning technology, and the UPT module can transfer user preferences in other scenarios. Besides, the M3REC framework uses a specifically designed backbone network together with a task-specific aggregate gate to promote all tasks to achieve the purpose of optimizing multiple tasks in multiple business scenarios within one model. Experiments on two public datasets have shown that M3REC outperforms those compared MSR and MTR state-of-the-art methods.

Memetic multi-agent system (MeMAS) is recently proposed as an enhanced version that integrates meme concept into multi-agent system (MAS) wherein all meme-inspired agents have an improvement in learning performance via meme evolution independently or social interaction. In the process of solving the black box optimization problem, the potential advantages of MeMAS have not been utilized well, which makes it a fertile area for further exploration. This paper presents a memetic multi-agent optimization paradigm through coordinate rotation (MeMAO-R) to combine MeMAS with evolutionary algorithms (EAs) to improve optimization efficiency. Based on MeMAS, the particular interest of MeMAO-R is placed on assisting original complex optimization task with new tasks generated by coordinate rotation. Further, MeMAO-R constructs the social interaction mechanism which facilitates to improve their convergence speed for solving the target optimization problem by utilizing meaningful information transferred across multiple agents with differing views of the target problem. Besides, MeMAO-R employs one or more classical EAs as the fundamental population based evolutionary solvers for multiple agents to optimize multiple tasks in a multi-agent scenario. Lastly, to testify the efficacy of the proposed MeMAO-R, comprehensive empirical studies on basic optimization problems are provided.

Artificial bee colony (ABC) algorithm has attracted growing interest for the continuous global optimization problems (CGOPs), where numerous algorithmic extensions have been developed. However, existing studies generally employ identical population size to perform the comparison among different ABC variants, regardless a fact that the generally suitable population size should be algorithm-dependent. Here we focus on the analysis of population size. This study is conducted in several well-known ABC variants under a set of benchmark CGOPs. We demonstrate that i) with the independently optimal population size, standard ABC can perform competitively comparing with its advanced variants, and ii) the most remunerative population size is related to the algorithmic exploitation/exploration ability. We anticipate that this study will provide useful insights to guide the appropriate usage of ABC, as well as its further enhancements.

In this paper, we integrate variable neighborhood search (VNS) into artificial bee colony (ABC) algorithm so that the search ability under variable neighborhood structures and local search is accelerated. Two VNS methods, including the reduced VNS and basic VNS, are integrated to develop different versions. The proposed algorithm, named variable neighborhood ABC (VNABC), is verified in a comprehensive set of benchmark functions. The experimental results confirm that VNABC outperforms the state-of-the-art ABC and differential evolution (DE) algorithms in terms of the convergence speed.

As a relatively new population-based technique for the continuous optimization problems, artificial bee colony algorithm (ABC) has verified its superiority and robustness over traditional evolutionary algorithms. However, regarding the wide variety of ABC extensions, the use of search experience produced during the search progress has been relatively unexplored to improve the performance. This results in an enormous waste of search resource to lower the search efficiency. In this paper, we develop a search experience-based search adaptation (SESA) approach to accelerate the search efficiency of ABC by taking three search resources into account, that is, the successful search directions, the fitness resource and the large number of failed solutions. The experiments over a comprehensive set of benchmark functions demonstrate that SESA can significantly accelerate the search efficiency of ABC.

Traditional accuracy-based XCS classifier system generally learns and evolves classifiers from scratch when facing each particular problem. Inspired by humans with the ability to learn new skills by inducing knowledge from related problems, transfer learning (TL) focuses on leveraging the knowledge of source domains to help the problem solving of another different but related domain. This paper attempts to combine XCS and TL to propose a novel extension transferable XCS (tXCS). tXCS utilizes the inherent characteristics of XCS, that naturally discovers expressive classifiers as the generalized knowledge of domains, to realize the classifier transfer from source domains to a target domain that makes it learn faster, which is conceptually different from the previous integrations between XCS and TL. The systematic study is presented to verify the ability of knowledge transfer between domains with different degrees of similarity, which has been pointed out to be the challenge of TL. We demonstrate that tXCS can significantly speed up the learning efficiency of canonical XCS in both of single-step and multi-step benchmark problems.

Though there are numerous approaches developed currently, exploring the practical applications of estimation of distribution algorithm (EDA) has been reported to be one of the most important challenges in this field. This paper is dedicated to extend EDA to solve one of the most active research problems - stock trading, which has been rarely revealed in the EDA literature. A recent proposed graph-based EDA called reinforced probabilistic model building genetic network programming (RPMBGNP) is investigated to create stock trading rules. With its distinguished directed graph-based individual structure and the reinforcement learning-based probabilistic modeling, we demonstrate the effectiveness of RPMBGNP for the stock trading task through real-market stock data, where much higher profits are obtained than traditional non-EDA models.

Traditional evolutionary algorithms (EAs) generally starts evolution from scratch, in other words, randomly. However, this is computationally consuming, and can easily cause the instability of evolution. In order to solve the above problems, this paper describes a new method to improve the evolution efficiency of a recently proposed graph-based EA - genetic network programming (GNP) - by introducing knowledge transfer ability. The basic concept of the proposed method, named GNP-KT, arises from two steps: First, it formulates the knowledge by discovering abstract decision-making rules from source domains in a learning classifier system (LCS) aspect; Second, the knowledge is adaptively reused as advice when applying GNP to a target domain. A reinforcement learning (RL)-based method is proposed to automatically transfer knowledge from source domain to target domain, which eventually allows GNP-KT to result in better initial performance and final fitness values. The experimental results in a real mobile robot control problem confirm the superiority of GNP-KT over traditional methods.

Accuracy-based XCS classifier system has been shown to evolve classifiers with accurate and maximally general characteristics. XCS generally represents its classifiers with binary conditions encoded in a ternary alphabet, i.e., {0,1, #}, where # is a “don't care” symbol, which can match with 0 and 1 in inputs. This provides one of the foundations to make XCS evolve an optimal population of classifiers, where each classifier has the possibility to cover a set of perceptions. However, when performing XCS to solve the multi-step problems, i.e., maze control problems, the classifiers only allow the agent to perceive its surrounding environments without the direction information, which are contrary to our human perception. This paper develops an extension of XCS by introducing cyclic conditions to represent the classifiers. The proposed system, named generalized XCS classifier system (GXCS), is dedicated to modify the forms of the classifiers from chains to cycles, which allows them to match with more adjacent environments perceived by the agent from different directions. Accordingly, a more compact population of classifiers can be evolved to perform the generalization feature of GXCS. As a first step of this research, GXCS has been tested on the benchmark maze control problems in which the agent can perceive its 8 surrounding cells. It is confirmed that GXCS can evolve the classifiers with cyclic conditions to successfully solve the problems as XCS, but with much smaller population size.

Genetic Network Programming (GNP) is derived from Genetic Algorithm (GA) and Genetic Programming (GP), which applies evolution theory to evolve a population of directed graph to model complex systems. It has been shown that GNP can solve typical control problems, as well as many real-world problems. However, studying GNP is mainly focused on the specific aspect, while the fundamental characteristics that ensure the success of GNP are rarely investigated in the previous research. This paper reveals an important feature of GNP - reusability of nodes - to efficiently identify and formulate the building blocks of evolution. Accordingly, adaptive GNP is developed which self-adapts both crossover and mutation probabilities of each search variable to circumstances. The adaptation allows the automatic adjustment of evolution bias toward the frequently reused nodes in high-quality individuals. The adaptive GNP is compared with traditional GNP in a benchmark control testbed to evaluate its superiority.

Artificial bee colony (ABC) algorithm is a relatively new optimization technique that simulates the intelligent foraging behavior of honey bee swarms. It has been applied to several optimization domains to show its efficient evolution ability. In this paper, ABC algorithm is applied for the first time to evolve a directed graph chromosome structure, which derived from a recent graph-based evolutionary algorithm called genetic network programming (GNP). Consequently, it is explored to new application domains which can be efficiently modeled by the directed graph of GNP. In this work, a problem of controlling the agents's behavior under a wellknown benchmark testbed called Tileworld are solved using the ABC-based evolution strategy. Its performance is compared with several very well-known methods for evolving computer programs, including standard GNP with crossover/mutation, genetic programming (GP) and reinforcement learning (RL).

In recent years, numerous studies have drawn the success of estimation of distribution algorithms (EDAs) to avoid the frequent breakage of building blocks of the conventional stochastic genetic operators-based evolutionary algorithms (EAs). In this paper, a novel graph-based EDA called probabilistic model building genetic network programming (PMBGNP) is proposed. Using the distinguished graph (network) structure of a graph-based EA called genetic network programming (GNP), PMBGNP ensures higher expression ability than the conventional EDAs to solve some specific problems. Furthermore, an extended algorithm called reinforced PMBGNP is proposed to combine PMBGNP and reinforcement learning to enhance the performance in terms of fitness values, search speed, and reliability. The proposed algorithms are applied to solve the problems of controlling the agents' behavior. Two problems are selected to demonstrate the effectiveness of the proposed algorithms, including the benchmark one, i.e., the Tileworld system, and a real mobile robot control.

Recent advances in Learning Classifier Systems (LCSs) have shown their sequential decision-making ability with a generalization property. In this paper, a novel LCS named extended rule-based Genetic Network Programming (XrGNP) is proposed. Different from most of the current LCSs, the rules are represented and discovered through a graph-based evolutionary algorithm GNP, which consequently has the distinct expression ability to model and evolve the decision-making rules. XrGNP is described in details in which its unique features are explicitly mapped. Experiments on benchmark and real-world multi-step problems demonstrate the effectiveness of XrGNP.

A novel graph-based Estimation of Distribution Algorithm (EDA) named Probabilistic Model Building Genetic Network Programming (PMBGNP) has been proposed. Inspired by classical EDAs, PMBGNP memorizes the current best individuals and uses them to estimate a distribution for the generation of the new population. However, PMBGNP can evolve compact programs by representing its solutions as graph structures. Therefore, it can solve a range of problems different from conventional ones in EDA literature, such as data mining and Reinforcement Learning (RL) problems. This paper extends PMBGNP from discrete to continuous search space, which is named PMBGNP-AC. Besides evolving the node connections to determine the optimal graph structures using conventional PMBGNP, Gaussian distribution is used for the distribution of continuous variables of nodes. The mean value μ and standard deviation σ are constructed like those of classical continuous Population-based incremental learning (PBILc). However, a RL technique, i.e., Actor-Critic (AC), is designed to update the parameters (μ and σ). AC allows us to calculate the Temporal-Difference (TD) error to evaluate whether the selection of the continuous value is better or worse than expected. This scalar reinforcement signal can decide whether the tendency to select this continuous value should be strengthened or weakened, allowing us to determine the shape of the probability density functions of the Gaussian distribution. The proposed algorithm is applied to a RL problem, i.e., autonomous robot control, where the robot's wheel speeds and sensor values are continuous. The experimental results show the superiority of PMBGNP-AC comparing with the conventional algorithms.

This paper presents an algorithm to discover and reuse subroutines in Variable Size Genetic Network Programming (GNPvs) called Subroutine embedded GNPvs (SGNPvs). GNPvs is a general type of GNP, which has a direct graph representation with changeable size. In order to improve the performance of GNPvs, SGNPvs has been proposed, in which a subroutine mechanism has been introduced to GNPvs by module acquisition. In SGNPvs, useful subgraphs are extracted and reused for individuals. Through extracting new subroutines to replace the old subroutines, SGNPvs can evolve the subroutines as well as evolve the individuals. The simulation results verify the performance of SGNPvs on a well-known dynamic multi-agent test bed - Tileworld.

This paper proposed a novel EDA, where a directed graph network is used to represent its chromosome. In the proposed algorithm, a probabilistic model is constructed from the promising individuals of the current generation using reinforcement learning, and used to produce the new population. The node connection probability is studied to develop the probabilistic model, therefore pairwise interactions can be demonstrated to identify and recombine building blocks in the proposed algorithm. The proposed algorithm is applied to a problem of agent control, i.e., autonomous robot control. The experimental results show the superiority of the proposed algorithm comparing with the conventional algorithms.

This paper proposes a different type of Genetic Network Programming (GNP) Variable Size Genetic Network Programming (GNPvs) with Binomial Distribution. In contrast to the individuals with fixed size in Standard GNP, GNPvs will change the size of the individuals and obtain the optimal size of them during evolution. The proposed method defines a new type of crossover to implement the new feature of GNP. The new crossover will select the number of nodes to move from each parent GNP to another parent GNP. The probability of selecting the number of nodes to move satisfies the binomial probability distribution. The proposed method can keep the effectiveness of crossover and improve the performance of GNP. In order to verify the performance of the proposed method, a well-known benchmark problem Tile-world is used in the simulations. The simulation results show the effectiveness of the proposed method.

Classical EDAs generally use truncation selection to estimate the distribution of the feasible (good) individuals while ignoring the infeasible (bad) ones. However, various research in EAs reported that the infeasible individuals may affect and help the problem solving. This paper proposed a new method to use the infeasible individuals by studying the sub-structures rather than the entire individual structures to solve Reinforcement Learning (RL) problems, which generally factorize their entire solutions to the sequences of state-action pairs. This work was studied in a recent graph-based EDA named Probabilistic Model Building Genetic Network Programming (PMBGNP) which can solve RL problems successfully. The effectiveness of this work is verified in a RL problem, i.e., robot control, comparing with some other related work.

As an extension of Genetic Algorithm (GA) and Genetic Programming (GP), a new approach named Genetic Network Programming (GNP) has been proposed in the evolutionary computation field. GNP uses multiple reusable nodes to construct directed-graph structures to represent its solutions. Recently, many research has clarified that GNP can work well in data mining area. In this paper, a novel evolutionary paradigm named GNP with Estimation of Distribution Algorithms (GNP-EDAs) is proposed and used to solve traffic prediction problems using class association rule mining. In GNP-EDAs, a probabilistic model is constructed by estimating the probability distribution from the selected elite individuals of the previous generation to replace the conventional genetic operators, such as crossover and mutation. The probabilistic model is capable of enhancing the evolution to achieve the ultimate objective. In this paper, two methods are proposed based on extracting the probabilistic information on the node connections and node transitions of GNP-EDAs to construct the probabilistic model. A comparative study of the proposed paradigm and the conventional GNP is made to solve the traffic prediction problems using class association rule mining. The simulation results showed that GNP-EDAs can extract the class association rules more effectively, when the number of the candidate class association rules increases. And the classification accuracy of the proposed method shows good results in traffic prediction systems.

Time Related Association rule mining is a kind of sequence pattern mining for sequential databases. In this paper, a method of Generalized Association Rule Mining using Genetic Network Programming (GNP) with MBFP(Multi-Branch and Full-Pathes) processing mechanism has been introduced in order to find time related sequential rules more efficiently. GNP represents solutions as directed graph structures, thus has compact structure and partially observable Markov decision process. GNP has been applied to generate time related candidate association rules as a tool using the database consisting of a large number of time related attributes. The aim of this algorithm is to better handle association rule extraction from the databases in a variety of time-related applications, especially in the traffic volume prediction and its usage. The generalized algorithm which can find the important time related association rules has been proposed and experimental results are presented considering how to use the rules to predict the future traffic volume and also how to use the traffic prediction in the optimal search problem.

The optimal route search in car navigation Systems is often considered to be a route search from the origin to destination. Many algorithms have been proposed to search for the optimal route from the origin to destination. However, in real situations several restrictions may need to be considered in the route search like some intersections must be included in the route while some should be excluded. The conventional optimal route search methods cannot consider such restrictions in the route search. In this paper, we propose a method to find the optimal route considering such restrictions, focusing on the restriction that some intermediate destinations must be visited before reaching the final destination. The proposed method is divided into three steps. In the first step, the optimal traveling times among the origin, intermediate destinations and final destination are calculated. In the second step, the optimal order of visiting intermediate destinations is optimized using RasID-D, a random search method for discrete optimization problems. Finally, in the third step, the optimal route from the origin to destination via intermediate destinations is determined. The paper also discusses the heuristic initialization to increase the efficiency of the optimal search. The proposed method was evaluated using a grid network with randomly generated intermediate intersections. Simulation results showed that the proposed method is more efficient than the genetic algorithm for optimizing the visiting order.