Collaborative Systems (U of Washington CSE Colloquium 2004)

So, I was hoping that this talk would be right up our alley.... how do we help Scientists work with each other and work with computing resources to get science done.

Instead, this talk was mostly theoretical. It talked a little bit about the small collaboration of one person and one machine. The speaker was part of a project at Harvard called Writer's Aid. That was a simple text editor sort of deal that assumed you were just typing raw text for your paper. (From the screenshots, it may actually have let you use LaTeX or something, but it didn't appear to be WYSIWYG at all.)

Then, rather than having to remember what entry in your bibliography you are trying to cite, you type something like: [CITE:scott+brown&2001&low-light]. Then, while you continue typing your paper, the system thumbs through your BibTeX file. At the same time, it also browses CiteSeer^? and other such resources looking for papers that match those criterion.

Then, when you have a moment, you can walk through the list of CITEs you didn't specify yourself and pick from the list of references it found. If you pick one that wasn't already in your BibTeX, it adds it for you.

Then, the talk just went into psychology experiments. It was concerned mostly with how group intentions are modified over time. Then, they did some experiments about how individual goals relate to group goals. But, the experiments they did (or at least the ones she talked about), didn't seem to have any group goals.

The talk was largely concerned with understanding how individual actors may have goals that conflict with the group goal.

The only thing that I really came away from this feeling we might need to address is this... suppose we're trying to schedule a pipeline.... we need several elements to coordinate and run simultaneously. The way her talk went indicates that auctioning is the best (or at least the most studied) way to figure out which exact resources are going to do which exact things.

The idea would be... we want to schedule a pipeline. We advertise the work we need to get done. Resources on the net which could actually do the work make bids on what parts they can do and how fast they can do it and how soon they can start. Then, the auctioneer coordinates who will get these "subcontracts" for the pipeline job.

Now, what happens when resource A is taking longer than it was contracted to do and now resource B is being put behind schedule... and it's already made other contracts for what it will do after this contract? The results of her research seemed to be "this will happen" without much idea of what to do then.

Obviously, rebidding for outstanding portions would be desirable rather than bailing and having to start from square one. So, it may be something to think about in the architecture is how to form "contracts" with resources, how to amend those contracts, and how to get the systems to learn rather than consistently overbid or underbid.

-- PatrickStein - 31 Mar 2005

Perhaps each advertiser could accumulate a "reputation" based on the accuracy of past predictions. Those with better accuracy have a better reputation and their estimate is given more respect. Each bidder can access the reputation and decide on how to respond -- probably by raising the "price" for higher risk advertisers. Reputations would be built by getting reports back from bidders after job completion. Bidders could also have reputations on whether they actually completed jobs, etc.

-- HarveyRhody - 01 Apr 2005

Yes, it's probably preferable to have independent tracking of how accurate predictions are instead of depending on each entity to predict on its own. And, it is probably preferable to have this independent system notified by the auctioneer as opposed to the bidders. The auctioneer has an incentive to report. The auctioneer has an incentive to make sure that next time it needs to auction, it has real information about the bids.

-- PatrickStein - 04 Apr 2005