v0.2.1 -- Bugfix: revenue_per_conversion in policy_info xlsx output

Bugfix

• Revenue_per_conversion in policy_info xlsx output was often smaller than what it is actually.
• Info given to policies was also affected. This may have affected policies to perform suboptimally.
• Cause: mean(revenues) is not the same as sum(revenues)/sum(conversions) when revenues contains 0.0 when some conversions are zero.
• Now the correct information is computed, provided to policies, and printed to xlsx.
  • Thanks to pbchen for bug reporting

Ad-click simulator v0.2.0 released

v0.2.0 is here.

This update introduces significant changes in the simulator. Some policies may experience incompatibilities. Please read on.

New feature

Finer bidding space

• Now bids can take values from 0.0 to 10.0 in increments of 0.1
  • Bid values are now in float instead of integers. This may lead to some incompatibility issues in policies.

Multiple ad slot placement

• More than one ad slots can be placed. So non-highest-bidders now have chance to win clicks.
• The highest bidder gets the top slot, which is most visible and has the highest click probability
• Bidders with lower bid will get lower slots than bidders with higher bid.
• Tiebreaking takes place randomly
  • If all bidders bid the same, their ad placement order is randomly assigned.
  • Lower bidders will not be included in the tiebreaking.

Example: Bidders = [A,B,C,D,E,F] with bids = [3,4,5,5,4,3], with 3 ad slots.

• Bidders C, D has highest. Their ad may be placed on first or second slot. Each click will cost 4.
• Bidders B, E has next highest bid, but there is only one slot left. Either B or E may be placed on the third slot. Each click will cost 3.

New information available: winning_bid_avg

• this is average of actual bids from the bidders who got the clicks
  • this is different from winning_bid, which is the highest bid.
    • discrepancy is due to multiple ad slot placement, as bidders can get clicks without absolutely winning the auction with the highest winning bid.
  • reported per iteration, and provided to all policies.
  • also printed out in policy output xlsx files.

add number-of-impressions + simulator parameters are tuned

New feature

number of impressions

• The policy output now contains another column named num_impressions. It contains the number of impressions from which the clicks are generated.
• Actually, these values are approximates to keep simulator runtime low. Exact tracking for impression counts is a future task (needs rework in simulator loop structures).

time spent by simulator and policies

• A new output file, output_time_spent_in_seconds.xlsx , contains cumulative time spent for simulator and each policies.
  • Each row corresponds to each iteration.

Bugfixes

• Fixed unnecessary repeated line prints in some of policy output xlsx files, which may happen with last attribute tuple.
• Simulators run up to number of iterations defined by max_t parameter given in the init xlsx file. This max_t value is the same as the max_t provided to policies.

Simulator parameters tuned

• The auction generating parameters are modified to result in much more frequent auctions, clicks, and revenues to make statistical learning faster.
• The variance of revenue random variable is reduced to make statistical learning easier.
• The init file used for the first run is included (auction_ini_01.xlsx updated).