Swarm and Evolutionary Computation最新文献

Drones are increasingly utilized for transportation reconnaissance due to their expansive field of view, cost-effectiveness, and agility. They can pre-reconnaissance the traveling routes of the ground vehicles carrying valuable goods to ensure the safety of vehicles and their goods. This rises a novel routing problem for the Vehicles and their Pre-Reconnaissance Drones, which is an integrated point and arc routing problem with temporal and spatial synchronization. A mixed integer linear programming model is developed to formulate the problem with complex synchronization constraints. The Variable Neighborhood Search algorithm integrated with the Temporal & Spatial synchronization-based Greedy search and simulated annealing strategy is designed to solve the model. A practical case based on real urban data from Beijing, China, and random instances with different sizes are tested and compared with the proposed algorithm. Computational results indicated that the proposed algorithm can solve the problem efficiently and outperform the simulated annealing algorithm and the greedy algorithm.

Smart agriculture aligns with the principles of sustainable development, making it a crucial direction for the future agriculture. This study focuses on a cooperative plant protection task allocation problem (CPPTAP) of multiple unmanned aerial vehicles (UAVs) with a common deadline in smart agriculture. CPPTAP permits multiple UAVs to conduct pesticide spraying on the same field. The completion time for each task fluctuates due to the cooperation among UAVs. We present a mathematical model and learning-based memetic algorithm (L-MA) to maximize the total area of the fields to be sprayed. In the evolutionary stage, mutation and repair operators based on value information are applied to balance the exploration and exploitation, while a problem-specific local search strategy is designed to enhance exploitation capability. A knowledge-based UAV allocation method (KUAM) is employed to maximize UAV utilization efficiency and minimize conflicts. Throughout the search process, Q-learning is utilized to assist the aforementioned operators and make decisions on the number of cooperative UAVs on fields. The effectiveness of L-MA is validated by comparing it against other state-of-the-art algorithms. The results demonstrate that L-MA outperforms the compared algorithms at a considerable margin in a statistical sense.

The reliability-redundancy allocation problem (RRAP) is an optimization problem that maximizes system reliability under some constraints. In most studies on the RRAP, either active redundant components or cold standby components are used in a subsystem. This paper presents a new model for the RRAP of a system with a mixed redundancy strategy, in which all components can be heterogeneous. This formulation leads to a more precise solution for the problem; however, RRAP is an np-hard problem, and the new mixed heterogeneous model will be more complicated to solve. After formulating the issue, a novel design of an evolutionary strategy optimization algorithm is proposed to solve that. The problem consists of discrete and continuous variables, and different mutation strategies are designed for each. The new formulation of the problem and the new method for solving it lead to better results than those reported in other recent papers. We implement the new suggested heterogeneous model with the PSO and SPSO algorithms to better compare the proposed algorithm. Results show improvement in both system reliability and fitness evaluation count.

The increasing accessibility of extensive datasets has amplified the importance of extracting insights from high-dimensional data. However, the task of selecting relevant features in these high-dimensional spaces is made more difficult due to the curse of dimensionality. Although Evolutionary Algorithms (EAs) have shown promise in the literature for feature selection, creating EAs for high dimensions is still challenging. To address the problem of feature selection in high dimensions, a novel concept of Evolutionary Reinforced Markov Chain is proposed in this paper. The proposed work has the following contributions and merits: (i) The paradigms of evolutionary computation, reinforcement learning, and Markov chain are incorporated into an integrational framework for feature selection in high dimensional spaces in a recursive manner. (ii) To support the global convergence of the algorithm and manage its computational complexity, a restricted group of the most effective agents is maintained within the evolutionary population. (iii) The dynamic Markov chain process efficiently manages agent evolution and communication, ensuring effective navigation through the search space. (iv) Agents moving in the right way are rewarded with an increase in their associated transition probability, while the agents going in the wrong direction are discouraged with a decrease in their associated transition probabilities; this promotes the establishment of an equilibrium state and leads to convergence. (v) The effective size of successful agents is reduced recursively while progressing through different states to further facilitate the speed of convergence and decrease the number of features. (vi) The performance comparison with state-of-the-art feature selection methods shows a significant improvement and promise of the proposed method over the existing methods.

Compared with common multi-objective optimization problems, constrained multi-objective optimization problems demand additional consideration of the treatment of constraints. Recently, many constrained multi-objective evolutionary algorithms have been presented to reconcile the relationship between constraint satisfaction and objective optimization. Notably, evolutionary multi-task mechanisms have also been exploited in solving constrained multi-objective problems frequently with remarkable outcomes. However, previous methods are not fully applicable to solving problems possessing all types of constraint landscapes and are only superior for a certain type of problem. Thus, in this paper, a novel dynamic-multi-task-assisted constrained multi-objective optimization algorithm, termed DTCMO, is proposed, and three dynamic tasks are involved. The main task approaches the constrained Pareto front by adding new constraints dynamically. Two auxiliary tasks are devoted to exploring the unconstrained Pareto front and the constrained Pareto front with dynamically changing constraint boundaries, respectively. In addition, the first auxiliary task stops the evolution automatically after reaching the unconstrained Pareto front, avoiding the waste of subsequent computational resources. A series of experiments are conducted with eight mainstream algorithms on five benchmark problems, and the results confirm the generality and superiority of DTCMO.

With the increasing demand for surgeries, surgery scheduling become an important problem in hospital management. Efficient surgery scheduling can enhance the optimal use of surgical resources, leading to high efficiency of surgery assignments. This work addresses surgery scheduling problems with surgical resources setup time. A mathematical model is established to describe the considered problems with the objective of minimizing the maximum completion time of the surgeries (makespan). Second, a modified artificial bee colony (ABC) algorithm is proposed, named QABC. Six local search operators are developed based on the characteristics of the problem, aiming to strengthen the local search capability of the algorithm. To further improve the performance of the algorithm, this study combines a Q-learning strategy with ABC algorithm. During each iteration of the algorithm, the Q-learning strategy is employed to guide the selection of search operators. Finally, the effectiveness of the local search operators and Q-learning based local search selection is verified by solving 20 cases with varying scales. And the results obtained by the Gurobi solver are compared with the proposed QABC. Furthermore, the proposed QABC is compared with the state-of-the-art algorithms. The experimental results and comparisons show that QABC is more effective than its peers for solving the surgery scheduling problems with setup time.

In this paper, we address the problem of dispatching multiple automated guided vehicles (AGVs) in an actual production workshop, aiming to minimize the transportation cost. To solve this problem, a self-adaptive memetic algorithm with Q-learning (Q-SAMA) is proposed. An improved nearest-neighbor task division heuristic is used for generating premium solutions. Additionally, a Q-learning is integrated to select appropriate neighborhood operators, thereby enhancing the algorithm's exploration ability. To prevent the algorithm from falling into a local optimum, the restart strategy is offered. In order to adapt Q-SAMA to different stages in the search process, the traditional crossover and mutation probabilities are no longer used. Instead, a self-adaptive probability is obtained based on the population's degree of concentration, and the sparsity relationship among individuals' fitness. Finally, experimental results validate the effectiveness of the proposed method. It is able to yield better results compared with other five state-of-the-art algorithms.

When evolutionary algorithms are employed to tackle constrained multiobjective optimization problems (CMOPs), constraint handling techniques (CHTs) play a pivotal role. To date, several CHTs have been designed, but they are only effective for certain types of constraint landscapes. For CMOPs with unknown properties, their optimization performance and efficiency remain uncertain. To tackle this issue, we attempt to mine and utilize the knowledge of constraint landscape to solve CMOPs. Specifically, the evolutionary process can be divided into three stages: learning stage, classification stage, and evolving stage. During the learning stage, the two populations, namely mainPop and auxPop, cooperatively evolve with and without considering constraints, respectively. The mainPop can locate the feasible regions, while the auxPop is employed to evaluate the size of the feasible regions. Subsequently, in the classification stage, based on the learned landscape knowledge, the category of problem can be determined: CMOP with small feasible regions or CMOP with large feasible regions. Then, in the evolving stage, for CMOPs with small feasible regions, CHTI, which includes a population exchange method and a feasible regions relaxation method, is proposed, while for CMOPs with large feasible regions, CHTII, which encompasses a dynamic resource allocation method and a coevolutionary method, is designed. The proposed framework is executed on extensive benchmark test suites. It has achieved superior or at least competitive performance compared with other state-of-the-art algorithms. Furthermore, the framework has been successfully implemented on the robotic manipulator path planning problem.

In the dynamic multi-objective optimization problems, if the environmental changes are detected, an appropriate response strategy be employed to respond quickly to the change. The predictive mechanism is effective in detecting the patterns of change in a problem and is often used to track the Pareto Frontier (PF) in a new environment. However, these methods often rely on the historical optimization results to approximate new environmental solutions, which can lead to back-predictions and mislead population convergence because of the low quality of historical solutions. This paper proposes a dual mechanism of prediction and archive (DMPA_DMOEA) to address the problem. The improvements include: (1) The well-distributed solutions from the previous environment be retained to ensure that reliable solutions exist in the new environment. (2) An LSTM neural network model is used to construct the predictor, which makes full use of the historical information and fits the nonlinear relationship between the pareto set (PS), thus improving the accuracy of the predicted solution. (3) These archived solutions and the predicted solutions collectively form the initial population for the new environment, which improves the quality of the initial population and maintains excellent tracking performance. Finally, Multiple benchmark problems and different variation types are tested to validate the effectiveness of the proposed algorithm. Experiment results show that the proposed algorithm can effectively handle DMOPs and has shown its remarkable superiority in comparison with state-of-the-art algorithms.

The successful deployment of Deep learning in several challenging tasks has been translated into complex control problems from different domains through Deep Reinforcement Learning (DRL). Although DRL has been extensively formulated and solved as single-objective problems, nearly all real-world RL problems often feature two or more conflicting objectives, where the goal is to obtain a high-quality and diverse set of optimal policies for different objective preferences. Consequently, the development of Multi-Objective Deep Reinforcement Learning (MODRL) algorithms has gained a lot of traction in the literature. Generally, Evolutionary Algorithms (EAs) have been demonstrated to be scalable alternatives to the classical DRL paradigms when formulated as an optimization problem. Hence it is reasonable to employ Multi-objective Evolutionary Algorithms (MOEAs) to handle MODRL tasks. However, there are several factors constraining the progress of research along this line: first, there is a lack of a general problem formulation of MODRL tasks from an optimization perspective; second, there exist several challenges in performing benchmark assessments of MOEAs for MODRL problems. To overcome these limitations: (i) we present a formulation of MODRL tasks as general multi-objective optimization problems and analyze their complex characteristics from an optimization perspective; (ii) we present an end-to-end framework, termed DRLXBench, to generate MODRL benchmark test problems for seamless running of MOEAs (iii) we propose a test suite comprising of 12 MODRL problems with different characteristics such as many-objectives, degenerated Pareto fronts, concave and convex optimization problems, etc. (iv) Finally, we present and discuss baseline results on the proposed test problems using seven representative MOEAs.