The idea we came up with was to modify our environment such that when we predict which bonds to add with the GCPN, it will generate additional graphs using the waters that are not yet participating in the clusters. Then, the GCPN will be allowed to create moves which bond the current cluster to these new graphs in the non-participating waters.