Note: This essay was written in 2010 as part of a course on advanced bio-inspired methods at the Norwegian University of Science and Technology (NTNU).
Introduction
The field of bio-inspired artificial intelligence approaches the problem of creating intelligent systems from a bottom-up perspective inspired by systems and processes seen in nature. Through the study of biology we have learnt that many complex natural systems are made up of relatively simple parts, or are the results of relatively simple processes. Through his theory of natural selection, Charles Darwin showed how complicated organisms can result from random variation coupled with non-random selection – showing us how increasing complexity can arise through a simple process. Other areas that have been studied, such as swarm behaviour in social insects and the workings of the nervous system, have also given us insights into systems where interaction between simple parts lead to complex results.
In the light of Darwin’s theory of evolution through natural selection it is not surprising that even the most complex systems show these properties. Evolution has no foresight and a complex end product must either have entailed an advantage at every step, or be the result of interaction between existing parts.
Applications of biology principles in computer science and engineering include evolutionary algorithms (EAs), artificial neural networks (ANNs), cellular automata, and swarm simulation. Especially EAs and ANNs have been successfully employed in many areas. Examples include circuit board design, process control, and optimization problems.
Scientists that want to study these biological processes also use these bio-inspired systems as models they can learn from. These researchers are naturally interested in making sure their models are true to nature – i.e. that the models are biologically plausible. But what about computer scientists and engineers: Should biological realism also be of importance to those of us who use these techniques to solve computational problems? Can we really expect to outsmart nature’s design by deviating from biology?
What is meant by bio-plausible?
We can look at bio-inspired methods as the superset and bio-plausible as a smaller subset. The criterion for inclusion in the smaller group of bio-plausible techniques is necessarily a fuzzy one. This is because the amount an approach may deviate from biology while still being plausible is a subjective matter, but also because our knowledge about biology is incomplete.
I have chosen to define the bio-plausibility of a technique as the extent to which it adheres to the biological reality here on Earth as we understand it. I am going to argue that we generally should not focus on bio-plausibility if we are are implementing bio-inspired techniques on computers except in a limited sense I will discuss near the end.
I will explore the topic by looking at example systems that use bio-inspired techniques.
Open-endedness
In his 1993 paper on artificial life 1 Thomas S. Ray argues that we impose limits on ourselves and our bio-inspired technology because we have only seen the one instance of life we ourselves are apart of. The examples he use are artificial life (AL) simulations and what he call AL instantiations.
In talking AL instantiation, Ray writes about how an open-ended process of evolution in a computational medium will – if we don’t impose any non-inherent limits – adjust to the medium it is developed in. Ray argues that we should view computer systems as alternate worlds with different rules governing their “physical reality”. This idea is in many ways reminiscent of the search for “alien biology”, i.e. life in other parts of the universe. While the laws of physics are the same everywhere (as far as we know); life that has emerged independently from the life on earth will most likely be very different – even if a similar process of evolution probably shaped it.
Touching reality
The problem with a completely open-ended approach is that we often have an explicit goal we want to achieve with our use of these techniques. If we are optimizing a process control unit for use in a power plant we most certainly want it to follow our rules of physics!
A good example of a technique which is first and foremost inspired by a biological phenomenon is the CLONALG algorithm presented by de Castro and von Zuben in [2]. This algorithm draws inspiration from immunology, especially the way the adaptive part of the vertebrate immune system learns to react to novel threats by proliferating antibodies that have successfully matched an antigen – in a process called clonal selection. CLONALG uses this concept by maintaining a pool of pattern matchers and cloning the ones that perform well. The algorithm has been successfully applied to optimization and pattern recognition tasks – the latter application being highlighted by de Castro and von Zuben in their paper.
de Castro and von Zuben emphasize that their algorithm does not accurately model a biological immune system, but merely draws inspiration from its high-level concepts. One factor they explicitly mention is left out cell–cell interaction.
CLONALG is thus clearly a bio-inspired algorithm. The questions it then: Would the algorithm perform better if more effort were put into making it more bio-plausible? The authors does not explicitly discuss this, but I believe the answer is no. When implementing an algorithm such as CLONALG for a specific usage area, the programmer will need to make various decisions in order to make the problem tractable. One will not be able to match the number of “cells” found in the immune system nor the infinite parallelism resulting from these cells’ independence. However, this is not necessarily a weakness; by familiarity with the problem, the programmer will be able to take the best “shortcuts”. Will the program often see identical objects? “Cheat” by storing the exact signatures. Will it rarely see the same item twice? Check if reducing the importance of “memory” helps.
Evolutions has tuned nature’s “algorithms” to their specific usage areas; if one wants to do something different we should in many cases expect it to be possible to get better results for different uses of the techniques.
Robustness and Adaptivity
A beneficial property of bio-inspired approaches is that they are robust and able to adapt to a changing environment. A system’s robustness is defined by Sipper in [3] as its ability to function in the face of faults, while adaptivity stems from the system’s ability to learn, evolve or self-organize, according to Sipper.
Tuning a technique to a specific area is often done at the expense of robustness and adaptivity. This is also relevant for the discussion about bio-plausibility vs bio-inspiration since one can argue that bio-plausible techniques are closer to what has been shaped by evolution for millions of years and are in that regard “proven” to work.
While this argument holds for problem areas where we are faced with noisy data and the need to change over time, most areas does not have a high degree of both of these factors. We are able to “outsmart” nature, but only because we pick our own battles, and – as Ray could have said: because a computer program does not occupy the same universe as biological organisms.
Science vs Technology
I have argued for a pragmatic, problem-solving – perhaps even engineering-centred – approach where we in my opinion should not focus on whether our algorithms and techniques are bio-plausible or not. However, bio-plausibility can be very useful in an indirect – but very important – sense.
In talking about CLONALG I mentioned that many areas of the immune system are not well understood and it is not unlikely that learning more about immunology will give us new insights and inspire new technology. In studies of the immune system and other biological processes bio-plausible models can be of great help in understanding these systems. First as simplified models and perhaps later as more complete simulations of the process being studied.
Artificial development, as described in [4], is an example of a field which has not yet had its breakthrough – at least in terms of being used to solve problems in practice. I believe that bio-plausibility is also a good “default” for exploring fields that are new or that are not yet well understood.
Insights into new areas will shed light on new inventions in nature and – just as today – people will be inspired to create technology exploiting these ideas, continuing the interplay between science and technology.