Memory Systems

[VKR]

It appears that to get this simulation to the next level and to get cohesive societies, an LSTM (Long-Short Term Memory) model will have to be implemented into the brains of these creatures. Perhaps they have some assigned memories decided by a gene and they inherit some from their parents, but the rest are "created" through experience. The wiki on LSTM neural networks shows more detail on this topic. Any thoughts? Thanks!

[Julian Hershey]

Hello!

We've noticed that the ScriptBots sim doesn't have very well developed memory features available to the bots, but we have noticed that it is possible for a box (neuron) to connect to itself, and if it gets switched on by anything else, it can stay on (or off), but such constructs are rare and mostly happen by chance. I've reviewed the topic, and checked out that wiki article you pointed out, and I think I might have an idea. Adding a third type of connection synapse, a kind of memory freeze input. Quite simply, if just one of these is greater than 50% active, then the box they lead to skips processing, effectively preserving its state. (slightly increased program efficiency is also gained). The next step would be for the agent to have their eyes and other senses connected to these memory-capable boxes and have a comparison neuron somewhere down the line connected to the sense and the memory box, but I for one don't want to interfere with the "natural" selection process of finding what works. If this type of memory can develop and is better for survival, we should know after I run some trials.

[VKR]

Hey I'm checking back REALLY late on the reply but how did the testing of this go? And I have some new suggestions for memory modeling. I've been learning more about deep belief/learning networks over the past few months on my own.