Transfer learning in multi-agent systems through parallel transfer

Transfer Learning(TL) has been shown to significantly accelerate the convergence of a reinforcement learning process. TL is the process of reusing learned information across tasks. Information is shared between a source and a target task. Previous work has required that the target task wait until the source task has finished learning before transferring information. The execution of the source task prior to the target task considerably extends the time required for the target task to complete. This paper proposes a novel approach allowing both source and target task to learn in parallel. This allows the transfer to be bi-directional, so processes can act as both source and target simultaneously. This, in consequence, allows tasks to learn from each other's experiences and thereby reduces the learning time required. The proposed ap- proach is evaluated on a multi-agent smart- grid scenario.