<p>The prototype reduction methods, as an important data pre-processing task, can improve instance-based classifiers by removing suspicious noise and/or redundant samples. Recently, a series of traditional non-evolutionary or evolutionary prototype reduction methods with traditional heuristic strategies or evolutionary algorithms have been developed. Despite having shown competitive performance, they still suffer from the following issues: (a) almost all traditional non-evolutionary prototype reduction methods heavily rely on specific assumptions about geometric and class information, leading to low robustness; (b) most evolutionary prototype reduction methods may face great challenges on search efficiency with increase in the sample number of the training set, leading to difficulties in extension and application; and (c) most evolutionary prototype reduction methods need to set too many parameters, leading to unstable performance. To advance the state-of-the-art prototype reduction methods by overcoming the above issues, a novel prototype reduction method based on an accelerated binary particle swarm (PRAB3PSO) is proposed. The main novelties are in the following. First, a new accelerated binary bare-bone particle swarm optimization (AB3PSO) with a new search space reduction strategy and a new particle update mechanism is proposed. Second, AB3PSO is used to select a global optimal reduced set from the original training set with no assumptions and few parameters. Intensive experiments have proven that (a) PRAB3PSO outperforms eight state-of-the-art non-evolutionary or evolutionary prototype reduction methods on UCI datasets from extensive industrial applications in weighing the instance reduction rate and classification accuracy of three instance-based classifiers and (b) PRAB3PSO is faster than state-of-the-art evolutionary prototype reduction methods.</p>