Multi agent systems and communication.
Agents may develop emotional contexts that contain information about meeting places with other agents (because of the possibility of long term rewards).
It would be a tough balance to find the situation in which this occurs, because if the environment is too large, they’ll never make it back to the same place, and if the environment is too small then they’ll be crossing over too much to need a meeting place. However, it is likely that there is a range of sizes in which this phenomena may occur.
The representation of contingency tables and emotional contexts should be very seriously considered because of the potential effects on the results.
A possible extension to this experiment would be the additional comparison of genetic programs. Using a stack-based genetic programming language, it would be very easy to make a direct comparison between the GP system and the NN system. The way one would do this would be to have the exact same inputs to both, and place all of the inputs on the stacks for the GPs, and have no more and no fewer input neurons. This would lend itself to a direct comparison of a genetic program based brain and a neural network based brain.
The algebraic formalization could be considered to be a sort of control, possibly adding in a purely random agent, maybe even a purely reactive agent as controls as well. The genetic algorithm combination with contingency tables and emotional contexts would then be sort of a stunted genetic program system.
The inputs to the agents should also be seriously considered, because of my personal experience with the effect this will have on the resulting actions of the agents, especially in a genetic algorithm context.
It is almost viable to think about the emotional contexts as concepts, or at least containing concepts. If one thinks about an agent conveying the concept of a food around a corner based on a given path, that agent is actually conveying two ideas, one, the path an agent can take to get food, and two, the idea of a corner. However, the idea of a corner will not be accessed by the receiving agent, at least in the approach that is being considered in this experiment. This suggests that an interesting research question would be if all the agents convey emotional contexts about both positive and negative things, however no bias is attached to the emotional contexts, leaving the receiver to decide whether to listen to the other agent. There are many other things that can be considered in this idea as well.