Improving time series prediction with ensembles of integrated photonic reservoirs

The feasible size and computational power of integrated passive photonic reservoirs is limited, amongst others, by high optical losses and wiring effort. As a step towards resolution of these issues, we propose to combine several smaller reservoirs to match or exceed the performance of a single larger one. In this abstract we investigate two possible combination techniques and evaluate their performance using the well-known Santa Fe timeseries prediction task. Our findings suggest that an ensemble of several smaller reservoirs outperforms a single reservoir with the same amount of overall nodes.