chrisokasaki
commented
3 years ago Early on, I thought something like this would be a popular choice. Somebody on BoardGameGeek even ran a couple of experimental trades based on maximizing happiness rather than number of trades. But it never caught on. Instead, the community settled on maximizing number of trades, often with maximizing number of distinct users as a secondary criterion.
In fact, many trades abandoned priorities altogether, at least partly because managing those priorities became unwieldy as the sizes of trades grew, especially when those wantlists were created by third-party tools where the user was clicking on the items they wanted.
I think the fundamental issue is that most people would prefer to have a greater chance to get something on their list, even if it was the least wanted item on their list, than have a greater chance of getting nothing (even if that meant that if you did luck out, you might end up with something better than you otherwise would have).
tjhighley
mrbrdo
Hello,
Not sure if you still follow this project, but I have a possible PR with some changes that could improve the "happiness" of the choices with the penalty of number of trades. I am interested in your opinion if this would be useful in the real world. I already did the code modifications and came up with the following results, where avg. fulfilled want position means the average position of the received item on the receiver's wantlist (#1 being most wanted):
Meaning that originally the people that traded got their 15th most desired item on average, while with the updated algo they got their 10th most desired item on average. However the median - most often - received desired item only changed from number 8 to number 7 in this case.
Explanation of the thought process and changes:
The original algorithm's core goal is to maximize the number of trades, hence the weights of non-trades is set very high. In the case with no wishlist ranking (preferential ordering, let's call it wishlist ranking), this is probably the optimal way. However, when there is a wishlist ranking, depending on the input data, this can mean that many people will get the less desired items on their wishlist, because the algorithm values the number of trades much more than the wishlist ranking. By lowering the weight on the non-trade, it is possible to achieve better fulfilled wishlist rankings (people getting the items they want more), but the number of trades will be less. Of course, this then becomes a parameter question of where the trade-off is worth it. The parameters of this trade-off are, in my opinion, the percent of total trades willing to sacrifice, for the percentage of average wishlist ranking improvement. This could be done through iterative repetitions and bisection (if needed) to find the desired balance between total trades and best wishlist rankings. I only did a quick modification and test on the sample data to confirm this idea. What do you think about this? In the ask.txt sample, we got 35% a better fulfilled average wishlist rank for the cost of 12% less trades - 12% of people that would otherwise trade, did not trade in this case, but the ones that did trade got their 10th favorite item instead of their 15th favorite, on average.
Another thing I was thinking about is the weighting of the first and last item on the wishlist. I think generally, people will much prefer the first item in the wishlist, and at least in case of a shorter wishlist, much dislike the last item in the wishlist (still better then a no trade of course). So I also experimented with adjusting the weights of the first and last item in the wishlist, by simply adding a fixed penalty to non-first items and twice that for the last item. In the case of the same ask.txt example, with some value I tested I got 60% more people getting first-choices (16 instead of 10 people), at the expense of the average wishlist ranking going from 9.8 to 9.96 and getting 2 trades less (171 instead of 173).
My thinking is to create a simple parameter in the way of a small range of preference (e.g. number 1-5), which would determine how much trades to sacrifice for better average wishlist ranking (e.g. sacrifice up to 10% trades for at least 30% better rankings), and also how much trades and average ranking to sacrifice for getting more "number 1" picks and less "last picks".
Any opinion on this?