If you look at different social insect species like wasps, ants, termites, etc. they behave in strange ways. For example, ants are incredible at searching for food. Different ants communicate with each other through their environment by secreting something called ‘pheromones’ that guide other ants in their behaviour. Together, they devise incredible foraging strategies that has been widely studied by many. You now have ant colony optimization algorithms (ACO) that are probabilistic techniques inspired from these ants that are used to solve complex computational problems. Likewise, each social insect shows very interesting behaviour and it is worthwhile understanding them because of their potential applications to other fields. Something very interesting that I got myself into a few weeks ago is a project on nature-inspired construction guided by a professor from my college (Prof. Srinivasa Chakravarty). The idea revolved around understanding how social insects like wasps and termites efficiently build together interesting architectures for their nests and how we could try replicating that with robots in civil engineering. Can we learn their ways and also build complex interesting architectures? I spent a lot of time reading about the way social wasps do it and I wanted to share the idea in this post.
A typical wasp nest is composed of combs containing reproductive chambers called cells and the combs joined together with structures called pedicels. Some species also build an external covering called the envelope. The materials they use for construction are usually different plant fibres which they combine with their salovary secretions to form paper carton which is eventually used as building blocks. They also use other materials like mud, sometimes. The queen wasp first sets about to search for a potential place to build their wasp nest and once it finds a suitable one, it places one brick. Then, the worker wasps continue from there and their building process, choice of materials and resulting archiecture all depends on the environment and the species. The worker wasps in some way seem to know what to do. One of the first theories that tried explaining this was the blueprint theory proposed by Thorpe. He said that the worker wasps have some kind a blueprint of the nest it wants to build and that is what it uses to guide the process. However, this was proved to be wrong through a demonstration. You try to disturb the process in between, say, you remove some building blocks, if it were a fixed blueprint that the workers used, then there shouldn’t be any effect of the disturbance on the final architecture. But it turned out, the architecture became completely different. This demonstration resulted in a change of the idea. Perhaps, it is not a blueprint but rather the environment and some knowledge that guides their actions. One such model that attempted at this was introduced by Theraulaz and Bonabeau. They modelled the decision factor as the local neighborhood configuration of a worker wasp compared to a known knowledge of configurations that is really needed. There exists a ‘ruleset’ which consists of a lot of if-then rules based on the neighborhood configuration pattern that if matched with the current neighborhood should then provoke the wasp worker to take a certain action, like place a block of material type 1 or of type 2 or do nothing. If the current neighborhood does not satisfy any rule, the worker just moves on. The idea here is that it is not an exact blueprint that drives the building of these architectures but an abstract pattern based knowledge that result in the kind of architectures that are out there. He attempted at reproducing the natural architectures with different models having specific rulesets and the results were pretty close. He also investigated on the effect of rulesets. You can identify that certain rules correspond to certain patterns in the archiecture and changing them in some way results in an expected change in some parts of the final architecture. For example, you might have some rules corresponding to building a pedicel and tweaking them could result in thicker or longer pedicel structures.
Apart from thinking about how exactly the workers know what to build, it is also important to ask how do the workers collaborate and build the same thing? There should be some communication between them if it is not a fixed blueprint that each agent is trying to stick to. Do the wasps talk to each other all the time? How does a worker know about the things that have been done at some place by some other worker and what is left to do? Here is where an idea got inspired from ants. The ants when foraging do not really talk to each other all the time. Rather, they talk indirectly by leaving pheromones on the environment. If one ant sees pheromone trails at a place, it tends to follow it. Some ant which left the trail has indirectly communicated with this one saying “Hey, this is the path that I found. Join me”. Something like that. So, even the wasp workers should be talking to each other in a similar way. That is the only way you get coherant resultant architectures. Now, sure it is not pheromones but its something which people call ‘stigmurgy’. It is of two kinds- qualitative and quantitative and works in a similar way as to pheromones. The model that I had mentioned previously makes use of qualitative stigmurgy. The pattern based ruleset idea was reinforced by a thought that the local neighborhood pattern to a worker is in some way a message left by another worker. The pattern forms the stigmurgy and influences the workers decision. Now, there is no proof per se proving this. It might as well be quantitative stigmugy or no stigmurgy at all but something very different. But this idea was good enough to reproduce almost similar natural-like wasp nest architectures. So, it is widely accepted. In the end, the goal is not to accurately understand how wasps work but to try modelling it logically as best as possible for use in other applications.
Now, after reading all these papers, I find it really great as a model to replicate wasp behaviour. But to actualy get a complex archiecture similar to natural ones, it takes carefully crafted set of dozens of rules that humans study the nests and figure out. This is still fine. My interest, is however, to eliminate this need. The wasps were not told by someone that this is how you do it but rather they learnt it. These papers were published quite a while back. Today, reinforcement learning has reached great feats and it really helps in training agents in a trail-and-error fashion. This problem seems to fit this setting and I want to see what can I do with it. How do you deploy multiple agents onto an environment which learns to build natural-like archiectures from environment feedback? Rather than replicating the existing natural archiectures if you could find how this came up, what kind of feedback did the environment give and model that, then these agents can come up with new interesting complex architectures. And you also have freedom in tweaking the environment setting now to suit your needs.
This became the topic of interest in my project which I will be further exploring.
See ya!
Ref: https://www.sciencedirect.com/science/article/abs/pii/S002251938570255X