Evaluation of The Stochasticity in The Collective Bernoulli-ball System as A Physical Reservoir Computer

Abstract Reservoir computing is a type of machine learning approach that uses a fixed, randomly generated dynamic system called a reservoir to process input data for prediction and classification tasks. The performance of a reservoir computer is highly dependent on the properties of its physical reservoir, particularly its stochasticity. Currently, there is no practical guide for selecting an appropriate reservoir to accomplish desired tasks. In this paper, we utilized the collective Bernoulli-ball system as the reservoir and investigated the effect of stochasticity on its performance in a non-linear prediction task. Additionally, we examined other properties of the reservoir computer, such as the input delay, input gain, n-samples average, and prediction target delay. This paper reveals how stochasticity affects the prediction performance of a reservoir and provides some possible methods of adjusting the parameters mentioned above to mitigate the drawbacks of stochasticity. Our findings show that in low-noise cases the reservoir output follows the trend of the target, but sometimes jumps to a significant offset. For high-level noise, the system remains stable, but the trend information in the output is compressed since noises cover the effect of input, which finally makes the reservoir computer useless for the task.