On the other hand, in reinforcement learning, agents learn estimated values for their actions on their states. Their states generally consists of the observations of the other agents. One agent observe himself in a fixed place. But the others observe him in an arbitrary place. Then their state-spaces must be different to distinguish one agent from the others. Even if they learn the same task, the estimated values on different state-spaces are different. Hence they cannot share directly learned results with the others.
We propose here a method to transmit one agent's learned estimated values to the others. We assume that any two agents have at least one observation in common of the agents which constitutes their state-spaces. We also show an experimental result of applying this method to a pursuit problem in which multiple agents cooperate to capture a prey.