[Wecount] exploring Distributed AI

Jutta Treviranus jtreviranus at ocadu.ca
Wed May 20 03:14:33 UTC 2020

Hi Clayton,
Yes, all current systems are unable to to deal with diversity and complexity. All reduce diversity and complexity. The challenges of disability show the cracks and shortcomings in any decision system, combine that with the challenges of genuinely including the range of diverse abilities and perspectives in a distributed system and we see the flaws and shortcomings of any multi-agent network.

I love the research questions. I think it would be a great challenge for a collaboration with several DAI researchers.


On May 19, 2020, at 4:55 PM, Clayton H Lewis <clayton.lewis at colorado.edu<mailto:clayton.lewis at colorado.edu>> wrote:

This is all interesting work. As I think you are suggesting, there's an opportunity, and a need, to project a diversity perspective into this space. Your personal interactions may indicate a welcoming reception for this, which would be wonderful, but I wouldn't bet on it, from the papers. The work on reinforcement learning has the goal of obtaining robust convergence with limited implementation resources. The experiments in Yang et al. use homogeneous agents; hard to say about Mguni et al. In any case, neither paper discusses how the models behave with more or less diverse agent populations. This is methodologically understandable, but can be damaging. To caricature the position, "If we model what's really going on, it's too complicated for the model to work (or to produce consistent results across runs), so we impose simplifying assumptions."

An interesting way to pursue the opportunities here could be to engage with the group at Virginia Tech, to explore two questions. (1) Would they be interested in modeling the behavior of people with disabilities (or have they already done so)? (2) How would doing so affect the technical capabilities of their platform, as well as the predictions made? The background on (2) is, is their platform capable of modeling people with disabilities, or do platform limitations assume them out? It should be possible to obtain funding to explore this, if the Virginia Tech people are open to the idea.

In the reinforcement learning work it may also be possible to explore the question mentioned above: how do these mean field models respond to agent diversity?

The attached paper by Vesterby discusses relevant issues, I believe, in a different, though related, modeling domain, network systems. Some passages from it:

"These researchers have an interesting, well worked-out project. The dynamic model or formulation they have created clearly works—as a model or formulation. There are, however, two problems—first simplification, and second the inherent limitation of a mathematical model to adequately represent qualitative factors. Both problems result in a mismatch between what the analytical tools do and what real-world highly complex systems do.

"Reduction of the diversity of relations, while a simplification, does maintain a reduced but still real relation to real-world systems. Systems have components, and components influence one another. When the diversity of relations is reduced to simple weighted influence relations, there is still representation of the components and the overall pattern of their organization, and the overall pattern of their influence on each other. Thus, reduction to an influence diagram maintains some representation of complexity by having representations of the components and relations.

"This can work to some degree with the simpler (quantitative type) complex systems where all, or most, of the relations are of the same kind. But it does not work with a highly complex system, where the influences are of different kinds, because it loses the diversity of the components and the diversity of the relations, and thereby loses the roles diversity itself plays in the system, in particular the roles diversity plays in resilience. Due to the loss of diversity, this general network-based theoretical framework no longer represents the intrinsic nature of a highly complex system."

Cheers, Clayton

On May 18, 2020, at 7:13 PM, Jutta Treviranus <jtreviranus at ocadu.ca<mailto:jtreviranus at ocadu.ca>> wrote:

I may have written this too quickly and see I wasn’t very clear what it makes manifest. Deep learning makes manifest the risks of exclusion, ignoring the edges and weak signals, prediction based on probability, using data from the past, etc. … all of which are counter to inclusive design.

Similarly I heard from several DAI researchers that principles of collaboration, diversification of agents, organic self-organization etc. are better at solving complex problems in DAI. These are some of the things we have been exploring with  respect to human decision making and problem solving.

What triggered this line of thought was notes I wrote in the late 80’s when these articles came out:

Decker, K. S. (1987). “Distributed problem solving: A survey”. IEEE Transactions on Systems, Man, and Cybernetics, 17 (5): 729–740.

Durfee, E. H., Lesser, V. R., Corkill, D. D. (1989). “Trends in Cooperative Distributed Problem Solving”. IEEE Transactions on Knowledge and Data Engineering, 1 (1): 63–83.

I then started to explore where this had gone lately.

I found things like this:

Yeoh, W., Yokoo, M. (2012). “Distributed Problem Solving”. AI Magazine, 33 (3): 53–65

Mguni, D., Jennings, J., Munoz de Cote, E. (2018). “Decentralised Learning in Systems with Many, Many Strategic Agents”. arXiv:1803.05028v1.

Yang, Y., Luo, R., Li, M., Zhou, M., Zhang, W., Wang, J. (2018). “Mean Field Multi-Agent Reinforcement Learning”.arXiv:1802.05438<https://arxiv.org/abs/1802.05438>.

Waldrop, M. M. (2018). “Free Agents: Monumentally complex models are gaming out disaster scenarios with millions of simulated people”. Science, 360 (6385): 144–147.

Also this was interesting but not directly relevant:

Yang, X. S. (2009). “Firefly algorithms for multimodal optimization”. In: O. Watanabe, T. Zeugmann (Eds.), Stochastic Algorithms: Foundations and Applications, Vol. 5792 of Lecture Notes in Computer Science, Springer Berlin Heidelberg: 169–178.

I have a meeting with a group from UCLA who want to talk about applying inclusive design to DAI to address complex, edge problems.


On May 18, 2020, at 7:10 PM, Clayton H Lewis <clayton.lewis at colorado.edu<mailto:clayton.lewis at colorado.edu>> wrote:

That's an optimistic take, and encouraging to hear. I'd be interested in cites for some of the more open thinking you are seeing. What I've seen, which isn't much, works in a more formal framework, but I think you are right about what the bigger picture should be.

On May 18, 2020, at 8:25 AM, Jutta Treviranus <jtreviranus at ocadu.ca<mailto:jtreviranus at ocadu.ca>> wrote:

One of the things I noted when I began to explore AI again was that AI made manifest many of the topics we have explored in inclusive design in such a way that hard scientists could not dismiss inclusive design. One of the topics I’ve wanted to explore is Distributed Artificial Intelligence because it makes manifest the inclusion dimensions of the diversity and inclusion equilibrium. In Distributed Artificial Intelligence you need to construct the coordination, collaboration, competition or teamwork of multiple agents. You can work with small amounts of data. Design of a “social” system of sorts comes into play. DAI is also better able to handle non-linear, multi-variate learning spaces. Many of the DAI approaches are bottom up. One form, Agent Based Systems, is usually exploratory and descriptive rather than engineered and prescriptive. Approaches like Swarms self-organize and their robustness and self-repair functions happen because or in spite of the fact that they are decentralized and unsupervised.

There is even exploration of alternatives to majority rules decision making in this space.

The reason I want to explore these is not because of the potential applications but because of the way they can provide simulations of different forms of group organization and interaction that makes the points we have made manifest to hard scientists, engineers and others that do not trust the soft science of inclusive design.

I would love to hear your thoughts,


Clayton Lewis
Professor of Computer Science
Co-Director for Technology, Coleman Institute for Cognitive Disabilities
University of Colorado

Clayton Lewis
Professor of Computer Science
Co-Director for Technology, Coleman Institute for Cognitive Disabilities
University of Colorado


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.inclusivedesign.ca/pipermail/wecount/attachments/20200520/0060e2f0/attachment-0001.html>

More information about the Wecount mailing list