Kohonen's self-organizing feature maps (SOFM) may also be used to extract finite-state!behavior from the dynamics of a trained DTRNN. This has been done, among others, by Tino and Sajda (1995) . The neurons in a SOFM form a neuron field (NF); neurons are organized according to a predetermined topology and then these neurons are topologically mapped onto the state space of the DTRNN as follows: a position in state space is assigned to each one of the neurons in the NF in such a way that neighborhood is preserved: points that are close in DTRNN space are assigned to neurons in the NF that are close. Tino and Sajda (1995) use a star topology for the NF after assigning the points in DTRNN space to clusters, they determine intercluster transitions and determinize the transition diagram until transitions are compatible with a deterministic FSM. The resulting FSM is finally minimized.