Emergence of Functionality in Evolving Physical Networks

Self-assembling nonlinear wire networks in complex virtual environments have some amazing properties: They learn to play computer games such as Tetris, if the network inputs represent the state of the virtual environment and the network outputs control the environment with actuators and if the virtual environment offers energy as a reward for successful controls. The network dynamics appears complicated and irregular, but observables, which classify the parts according to their function in channeling the energy flow through the system, are quite reproducible. Experimental studies suggest that the tendency of networks to evolve towards self-organized critical states, states with large entropy production, and states with minimum resistance drives them to predict and control complex environments. We contrast the computational capacity of these networks with conventional digital computers.

Speaker: Alfred Hubler