One way to tackle the concept of ultrastability in evolved dynamical neural networks (to do  from previous post) is using the task generated by Ezequiel in his paper “Homeostatic adaptation to inversion of the visual field and other sensorimotor disruptions“.
There are several factors in his methodology any or all which could be playing a role in facillitating ultrastability. My interest is in finding out which (if not all) of the factors are relevant. The task is a simple light phototaxis one in a 2-dimensional agent and arena, but there are many interesting subtle variants in his approach:
- Application of noise (e.g. in sensor positions, in light intensity)
- Needing to actively search for the light (e.g. lights are very far away in combination with the previous noises).
- Keeping neurons active (i.e. not stable in some equilibrium)
- Keeping neurons out of saturated regions (i.e. light version of 3, equilibria allowed but not in some regions)
- Application of changes to the parameters of the evolved system when neurons cross certain thresholds. Still if 5 is necessary or facillitates homeostatic adaptation then we would like to know further whether we can replace Hebbian learning to the weighs to other:
- Rules (e.g. random changes?)
- Parameters (e.g. time-constants, biases?)
- Thresholds (time-based as opposed to absolute value e.g. too much time spent on a limit cycle? as opposed to neuron 3 went above 0.75.. ?)
All of these are currently confounded in the explanation of how agents described in Ezequiel’s paper adapt to sensorimotor disruptions.
The experiments that are needed (at least to start generating intuition) are simple. Evolve populations of agents to do phototaxis on each of the conditions (and combinations of) described above and test the ratio of success of agents that cope with visual inversion. I would really would like to find some spare time to do this!