ORIGINAL ARTICLE Escape velocity: a new operator for gravitational search algorithm U. Güvenç 1 & F. Katırcıoğlu 2 Received: 12 October 2016 /Accepted: 24 March 2017 # The Natural Computing Applications Forum 2017 Abstract Gravitational search algorithm (GSA) is based on the feature of reciprocal acceleration tendency of objects with masses. The total force, which is formed as an influence of other agents, is an important variable in the calculation of agent velocity. It has been determined that the total force and, thus, the velocity of the agents that are located far away, is low due to the distance. In this case, they continue their search in bad areas, as their velocity is low, which means a decrease in their contribution to optimization result. In this paper, a new operator called Bescape velocity^ has been pro- posed which is inspired by the real nature of GSA. It has been suggested that adding the escape velocity negatively will en- able the agents that remain far away or outside of group be- havior to be included in the group or to be increased in veloc- ity. Thus, the study of perfecting the herd or group approach within the search scope has been carried out. To evaluate the performance of our algorithm, we applied it to 23 standard benchmark functions and six composite test functions. Escape velocity gravitational search algorithm (EVGSA) has been compared with some well-known heuristic search algo- rithms such as GSA, genetic algorithm (GA), particle swarm optimization (PSO), and recently the new algorithm dragonfly algorithm (DA). Wilcoxon signed-rank tests were also utilized to execute statistical analysis of the results obtained by GSA and EVGSA. Standard and composite benchmark tables and Wilcoxon signed-rank test and visual results show that EVGSA is more powerful than other algorithms. Keywords Gravitational search algorithm . Escape velocity . Optimization algorithms 1 Introduction Gravitational search algorithm (GSA), which is carried out by the inspiration of the laws of gravity and mass interaction and is also a physics-based heuristic stochastic optimization algorithm, is first- ly presented by Rashedi et al. [1]. Beginning with the presentation, it has been analyzed by the researchers and it has created appro- priate solutions to complex problems and for industrial applica- tions and has shown good performance. However, the drawbacks such as early convergence or obtaining poor results have been observed especially in multimodal fitness functions and high- dimensional problems during the process of optimization, just as it takes place in some heuristic stochastic optimization algo- rithms. Various enhancement operators have been suggested within GSA in order to eliminate that disadvantage. In the study carried out by Sarafrazi et al. [2] in 2011, the agent with the best solution has been accepted as the star of the system and the other solutions have gained the characteristic of being dispersed and perished under the gravitational force of that star. The distance of the respective agent and its nearest neighboring agent star is controlled. The respective agent breaks down if it is smaller than the specific threshold value. While the iteration pro- ceeds, as a result of interim solutions, this approach has been proposed to prevent the complexity of algorithm from increasing and the agents from getting away from each other [2]. Within the GSA search procedure, the direction of the agent is calculated based on the total force that it receives from other agents. This situation has two disadvantages in the form of memory * U. Güvenç ugurguvenc@duzce.edu.tr 1 Faculty of Technology, Department of Electrical-Electronics Engineering, University of Duzce, Konuralp Campus, 81620 Duzce, Turkey 2 Duzce Vocational High School, University of Duzce, City Center Campus, 81100 Düzce, Turkey Neural Comput & Applic DOI 10.1007/s00521-017-2977-9