Monthly Archives: April 2013

Used Planet: A Global History

Are we living in the Anthropocene? Published Monday in the Proceedings of the National Academy of Sciences, USA, Erle Ellis and colleagues paint a picture of historical land-use that significantly shaped the Earth’s surface more than 3000 years ago.

See a blog post or media coverage in the New Scientist for more details and to download the paper.

What is an agent-based virtual laboratory?

I have recently had several conversations with colleagues and exchanges with reviewers about the exact meaning of the term ‘agent-based virtual laboratory.’ So, it seems like the perfect time to devote a single post entirely to unpacking this idea.

First, I should acknowledge that many researchers have used agent-based models (ABMs) of differing levels of detail or abstraction to explore the influence of particular processes or parameters on model outcomes. This is certainly a valuable exercise, and in this sense, the virtual laboratory approach is not new. However, these virtual lab efforts have been undertaken with site-specific, case-study ABMs – and this is where the distinction lies.

“All virtual labs can use ABMs, but not all ABMs are virtual labs.”

That is a favorite phrase I have used in multiple presentations on this topic. It is meant to make a clear distinction between ABMs as tools, and virtual labs as an approach. The type of virtual labs I use in my research and write about here are designed from the outset to be generalized modeling environments that can be applied across many different settings and locations. The advantage of this is that one can conduct comparative research – forming hypotheses of how and under what conditions certain processes or factors will be important or not. These hypotheses can be generated systematically across sites and then tested against empirical data. More conventional virtual lab efforts use case-specific ABMs and thus cannot be easily applied to different sites.

Figure4_rev (2)Take, for example, the figure to the right, which shows agricultural intensification patterns generated by agents with generalized decision models in response to alternative environmental conditions. With this framework, the influence of environmental suitability on land-use intensity – and its interactions with other processes, such as increased market influence and population pressure – can be experimentally explored across different sites.

With that said, if one desires to understand a particular context and/or predict possible future scenarios, the case-specific ABM approach is the way to go. There will always be a place for such models. But those models are not the best option for generating generalized knowledge and building theory. For that, a more generalized and transparent modeling framework is needed.

This approach is similar to that of “artificial systems research” that my friend and colleague Len Troncale describes. Quoting from one of his blog entries, a virtual lab approach “enable[s] adding or subtracting different sets of systems processes to see how these alterations effect sustainability of the resulting systems.” Thus, the goal of agent-based virtual laboratories is to explore and form testable hypotheses of how certain factors interact with agents’ decision-making processes to produce emergent system outcomes, and to do so across different land systems to build towards general land system theory.

Notes from AAG 2013

AAG2013Last week I attended the 2013 annual meeting of the Association of American Geographers in Los Angeles, CA. In particular, I attended the Land Systems Science Symposium and the Agent-Based and Cellular Automata Model for Geographical Systems sessions. It was great to catch-up with old friends and meet a few new colleagues. Now that the chaos of coming back to work after a week off has passed, I thought this would be a good time to reflect on the happenings of the conference.

Overall, I thought this was a much stronger meeting than last year’s. It was apparent from the Land Systems Science (LSS) Symposium that the field formerly known as “Land Change Science” is beginning to come into its own. The full scope of LSS was on display, as the first two days were dedicated to case studies from different world regions, followed on the last two days by LSS modeling, theory, and applications. I was also impressed by the agent-based modeling (ABM) sessions this year. I sense a real transition in ABM research, as the presentations demonstrated much deeper thinking about the implications of model result, and model building as an art and tool for learning. There also seemed to be a sense that ABM is no longer on the fringe – it’s no longer a new method and the ABM community can now discuss the challenges and weaknesses of ABM more comfortably. It was a sure sign that the ABM method and paradigm are maturing.

Several themes in particular caught my attention over the course of the week:

1. YAAWN – Yet Another Agent-based model … Whatever … Nevermind.

That clever acronym came courtesy of the organizers of the panel on ABMs and land-use change modeling organized by David O’Sullivan and Tom Evans. The general motivation for the title was the observation that the number for place-based, case-specific ABMs has exploded, and as a research community, it is worth asking, “What is the marginal gain from one more case-study ABM?” I, of course, was thrilled to hear such a question, as the drive for more systematic, generalized knowledge motivates my use of agent-based virtual laboratories. The question was posed to the panel, and I particularly liked Dan Brown‘s response. The message was that there will always be a role for case-study ABMs, but it is also necessary to balance the use of empirically detailed models with more abstracted models to build theory. This sentiment was reinforced by Sarah Metcalf, who argued that model hybridity was the next wave of ABM research.

2. Attempts to forge systematic ABM practice and knowledge.

Going along with the general observation of higher quality presentations, the thinking about ABM practice was notably deeper this year. I particularly enjoyed David O’Sullivan‘s presentation of creating a ‘pattern language’ for ABMs and cellular automata. The general idea is to create ‘building block’ models out of generalized processes/structures that facilitate the development of more complicated models. I found this idea analogous to Len Troncale‘s work with isomorphies and systems of systems processes theory. Another important question posed by James Millington was, “When should we use ABMs and how complex do they need to be?” Indeed, this is a fundamental question that should be revisited often.

3. Understanding model outcomes and variability.

Finally, another sign that the ABM field is maturing, there was much discussion about the importance of more thoroughly understanding uncertainty and variability in our models. Chris Bone presented an innovative temporal variant/invariant method for evaluating model performance, which shows much promise for deepening our understanding of the sources of variability within complex systems models.