Illustration: particle swarm optimization

It may be worthwhile to illustrate this discussion with example of Particle Swarm Optimization (PSO). A simple PSO algorithm can be imagined as a bunch of blind walkers (i.e. observers) parachuted on a landscape with irregular terrain (i.e. state-space). These walkers are collectively trying to find something (maybe the highest mountain, maybe the deepest valley - depends on their goal). Following the first approach (i.e. global state-space description), the information from all walkers gets aggregated by the global observer where the optimal decision is made and propagated "down" (this is an informal explanation of MapReduce algorithm). Now, there are many scenarios when information aggregation by a global observer is not possible or useful (prohibitively large data structures; lack of concentrated processing power or its vulnerability; undefined, changing, or simply non-existing global goals; and at last but least - constantly changing landscape of the terrain). In such case, global observer does not help. If mutual coordination is beneficial to the walkers, it has to emerge from the individual interactions between them, without central data aggregation.

The particularly interesting case in this illustration is constantly changing landscape where movements of particles have en effect on changes in the landscape. I think this case is most realistic in social domain - as a distributed system. We have been touching it from various perspectives when talking about scalable/stratified systems where stratas are related with all kind of feedback loops³⁴.