Extra study of equity

[akolonin]

Deborah wrote: "I would like to add a study of equity (which I would design) the purpose of which is to prove that utility increases as equity increases, in a more realistic simulation that has more realistic internal dynamics (Deborahs spec simualation) . It would deal with the heavy dependence of calculated goodness on market volume, which limits the ability of new agents without high initial market volume to enter the market. It would also show high Tonini Phi values as a result of the ability to grow." Note: if this is done, it has to be proved having expressively different values for different market conditions and different RS parameters, otherwise we can't use it.

[akolonin]

@deborahduong - can we sum this up, provide the poster for IJCAI and close then? Can we have reasonable variance of equity on adaptive-simulation to defend our IJCAI thesis?

[deborahduong]

Comment from Ben In Slack #reputation channel May 13 Ben Goertzel [2:29 PM] yeah sure no worries :wink: …. if it’s accepted, great! If it’s not accepted, I suggest to perhaps rework it so that the novel “modified Gini coefficient” measure is the main theme…. I think this measure is interesting in itself, assuming it’s original. The reputation system experiments then serve as an illustration of the measure in practice. This would reduce the critique some referees have had that there is no systematic comparison with other reputation systems. You’re putting a metric out there, showing how your system does by that metric, and inviting others to compete…

[deborahduong]

@deborahduong - can we sum this up, provide the poster for IJCAI and close then? Can we have reasonable variance of equity on adaptive-simulation to defend our IJCAI thesis? Sure I will get you the materials, but concerning closing the issue... I think this is an excellent use of my talents as applied to reputation systems in a way that benefits our team, as well as a central focus of the company, and so should be expanded instead of closed.

[deborahduong]

In the technical report, we did have an excellent result, that clearly reputation systems with greater equity also had greater utility. https://docs.google.com/document/d/1FEr4ir0jBZ5PZtOBas7naeHef8AnPo_E5IArSoxN0kE/edit This held across he board, and is publishable. However I also wanted to have a result to say that the way we use the reputation system matters, whether with winner take all or roulette wheel use matters, and we did not get that. I think the reason is that we did not follow the simulation best practice of taking stats when the system reaches a steady state (This practice was recommended earlier but was not prioritized). Leavers and enterers were added to the system in order to solve the problem of the initial randomness of choice of agents in the simulation causing a unrealistic result, that all were known and treated fairly. However the runs for the report were only 90 days, they did not surpass day 240 which would be the day in which the system reaches thermal equilibrium, that is, it forgets the initialization artifacts which shouldn't be included in the results. I believe short runs explain why the Winner Take All reputation system usage was only slightly less utility than the Roulette usage. When the Python system is ready, all reputation settings including the equity scenarios will be rerun in longer runs, to be put along with our other results in an AAMAS paper due Nov 12.