The film clips below are a demonstration of the ESP system applied to the double pole balancing problem. ESP evolves subpopulations of neurons that are used to form neural networks to solve the task. Neurons from each subpopulation are combined into networks that attempt to balance both poles without going off the track. Each generation the neurons that were used in forming the best networks are recombined to produce new neurons for the next generation. The film shows the behavior of the best network from each generation of an actual simulation. For this run, a network that could control the system for more than 30 minutes of simulated time was found in 12 generations.

NN Research Group
  Clips
Generations 1 - 9
In this first stage of evolution, the best networks can only balance the poles for at most 5 seconds. Initially, these networks are those that can delay failure by moving slowly in the direction of the falling poles. As the neurons begin to adapt, the network can at least swing the pole back to vertical.
MP4 Movie File 0.22 Mb
AVI Movie File 1.90 Mb
Generations 10 - 11
By generation 10, the networks show qualitative improvement. The poles are brought to vertical and some success is achieved in reducing the poles' angular velocity and damping oscillations.
MP4 Movie File 0.17 Mb
AVI Movie File 1.90 Mb
Generation 12
The task has been solved. The network is able to balance the pole for more that a 30 minutes.
MP4 Movie File 0.20 Mb
AVI Movie File 1.90 Mb
Generations 1-12
Full Clip
MP4 Movie File 0.62 Mb
AVI Movie File 5.27 Mb