Philosophy of Science and Computer Modeling

I have been reading Generative Social Science: Studies in Agent-Based Computational Modeling, by Joshua M. Epstein (2006). This book is helping me clarify my thinking about my project in this Perceived Order blog.

Here is one of my dilemmas. The principal model I use, being critters on a tabletop, is an agent-based model (ABM) which can be developed in one of two ways: either as a thought-experiment ABM or as a computerized ABM. I have worked at developing it both ways, but I had not given much consideration to the difference between these two ways until I started writing this blog. Is one way better than the other?

To present a thought-experiment ABM I need to be clear and convincing in my use of language and logic. Most of what I have presented so far on this blog has been in this thought-experiment vein.

On the other hand, to develop a computerized ABM, I need to develop computer code and experiment with running the model. In fact I have worked at this, perhaps two months worth of full time work. So let me say more about computerized agent-based modeling.

Con Computerized ABM

Computerized ABM does not promise much help in winning converts.

I have been involved with computerized ABM almost from the start, albeit in the periphery. In 1982, when I started a graduate program in computer science, I was excited by an idea which struck me as having great power. That idea was agent-based modeling, although that “ABM” label was not known to me at the time. I hoped to pursue that idea as a dissertation. I wrote and ran a few basic models at that time. During the thirty years since then I have revisited the literature and written models a few more times, because I continued to believe that this sort of modeling would become an important tool for social engineers and scientists.

But computerized ABM has not become a prominent tool in social science as I initially believed it would. I cannot think of any case in which news of accomplishments of computerized ABM has caught my attention, apart from those times when I intentionally sought such news.

Furthermore I suspect that others, possibly including Joshua Epstein, who became enthusiastic about computerized ABM two or three decades ago, also feel disappointment with the lackluster reception (though here I may be overreaching by reading between the lines). In spite of the charm imagined by many early enthusiasts, computerized ABM seems not to have made a big splash.

As such, the enthusiasm which I might feel for working to implement the Resource-Patterns Model of Life (RPM) in computerized ABM is diminished. My motive, after all, is to influence the perceptions of others, and not just to build a computer model of a social phenomenon.

Computerized ABM demonstrations are difficult to accomplish.

Joshua Epstein says:

In my own experience, given a macroscopic explanandum, it is challenging to devise any rules that suffice to generate it. (page 54, italics in the original)

I was gratified to read this because it validated my own experience. I have worked hard to find rules which suffice to generate even the simple phenomena which I believe can come from RPM. When I run the early draft of a model I am apt to see that what I am getting makes sense – after I think about it – but it is not what I had hoped and imagined I would get. So then I am apt to realize that I need to add a behavior-restraining parameter which I had not previously considered. The cycle repeats on subsequent runs: What I am getting makes sense, sort of, but it is not what I was aiming for. Often I decide to try tweaking the value of some parameter.

In spite of the months I have spent developing RPM as a computational ABM, I have not come close to achieving the line of trade (see the second picture on this post) which seems obvious and almost trivial in my thought-experiment mode. The creative effort requires attention to details which I had not at first imagined.

Then, after I have struggled and found some detailed specifications which lead to approximately-desired results, I am apt to believe those details are important. So I am apt to include those details in my written descriptions, even though those details may offer no help to a reader who is already struggling to understand my point. Many such details have found their way into what I have posted todate on this blog. But do those details help me write a convincing demonstration? I am skeptical.

Pro Computerized ABM

The education in unanticipated details may be valuable.

The effort to make a computer model generate an explanandum (How do you like my fancy new word?), even though it is difficult and unlikely to help with persuasion, can teach important insights about the explanandum. Computerized ABM is not empirical work, I suppose, but it seems more grounded in experience than mere thought experiments.

One important lesson stands out in my experience. I was taught a necessity in one of my first attempts to code a computerized ABM: An agent has to have a goal, something it is trying to accomplish. This did not become obvious to me until I had much of my code written for a model. I already had the agents and the actions available to each agent. The agents also had senses and memory. But when I started to write code to decide an agent’s next act – then and only then did it strike me that I had to code some purpose into the agent – or there was no basis for favoring one prospective act over another.

That lesson was forcefully implanted in my consciousness. But now, when I review the work of social scientists and commentators, I often get the impression that they are making the same mistake that I had made. They are overlooking and failing to mention the central role of the motive of an agent.

To Computerize or Not?

My effort in this blog to convey the Resource-Patterns Model of Life more successfully than I have in the past has challenged me to think carefully about what I am doing. I have been gaining new understanding of words such as: predict, explain, persuade, perceive, and show. Epstein adds another word to this list: generate.

His first three chapters deal with philosophy, with the challenges and limitations of computerized ABM. Evidently he has spent decades wrestling with questions such as I face now, and probably he has participated in many discussions on these questions. I still do not have enough experience to appreciate all that he conveys, but I am helped by his writing on the philosophical questions.

I will continue for the most part in only the mode of thought-experiment ABM, since I suppose that mode will be the most fruitful for my purpose of persuading others during the next year or so. But the deeper education offered by computerized ABM retains its appeal, especially for the geek in me who can enjoy immersion in code and experimentation.