SCML Second Data Release

You can download the dataset here.

This data release contains three files (one per track).

To avoid revealing any specifics about currently running agents, the dataset was processed in the following ways:

  1. All types and agent names are coded as T.... and A..... This is done in way that ensures that all instantiations of a given actual agent type are mapped to the same coded agent type. For example all agents of type DecentralizingAgent will be mapped to agents of type code say T234 (or something similar) in every world.
  2. Worlds from all tournaments are mingled together and their names are also randomized. This avoids the possibility of mapping agent type codes to actual agent type names based on their performance on different tournaments.
  3. Only some of the worlds of each tournament are included to avoid the possibility of mapping agent type codes to actual agent type names based on their aggregate performance.

It should be possible to visualize the logs of any world in the dataset using scml-vis. Note that you can compile all (or some) of these worlds together in a single database allowing you to visualize multiple simulations at once. For example, if you have extracted into the folder Oneshot, you can visualize all worlds in it at once (will be very slow) by running scmlv compile Oneshot followed by scmlv show -f Oneshot.

Detailed Content

The dataset contains the following set of files for each world simulation. You can find more information about them here:

File Description
params.json Basic world information
info.json Extended World information
agents.json Agent information
negotiations.csv Negotiation records (if any)
contracts.csv Contract records (Including exogenous contracts for std/collusion)
breaches.csv Breach records
stats.csv World and agent statistics