Create simple multi-agent bidder

[simofoti]

• [ ] Run multiple bidders simulation where bids are placed at the same time. Each bid in this scenario will be the output of our best performing single agent bidder, perturbed with random noise to simulate a stochastic game.

• [ ] Create a new dataset with updated bid and pay prices from the multiple agent simulation

• [ ] Re-train our best performing model on the new dataset

[simofoti]

OBS. the magnitude of the noise has to be defined. An idea might be:

• run our best performing bidder saving his bids (different set for bids he won and loose)
• observe distribution of differences between these two sets of bids and the payprice

This should provide us with an information of the amount of money I had to bid to win the auction with respect to the second highest bid (i.e. the bidprice if I won bidding against the dataset). I can assume that another agent would bid the same amount to outperform me.

Observing differences with respect to auctions I didn't win, I know that I should have bid more. If this distribution is in general with a smaller mean it could become a lower bound or something like that. If the mean is comparable or higher I could merge this info in the winning bid set.

Hoping in a well shaped distributions, I can fit the differences and sample from them a perturbation.

Here I could try to use different distributions fitted from data to simulate different unknown bidders behaviour.

These considerations should be than testes against arbitrary ("not-trained") noise.

[simofoti]

OBS 2 we might expect changes in the market price distribution after training the new model. Would be reasonable to assume that also other bidders are training their models with similar techniques and their marketplace estimates are different?

Instead of perturbing the bids an idea might be shifting the market prices distributions (most likely to higher values) and shrink/dilate the distributions.

Then compare the results with perturbations above-mentioned.