Datathon cases

Datathon2020 – Create gaming bot case – provided by Imperia Online

In the past few years DeepMind’s Alpha projects, IBM DeepBlue and OpenAI Five have shown that we are reaching the point of matching and exceeding human performance in complex game environments where there is no silver bullet to achieve a goal. In the push to mimic human performance in games Imperia Online has prepared a case in the game of Baloot. Have you ever heard about Baloot? Sounds like Belote and actually is like Belote but a little bit different.


Business overview

Anyone that has ever played digital multiplayer games, has experienced the frustration from waiting half an hour for a game to start just to have the game scraped because of somebody leaving in the middle of it. Well a nice solution to that would be an AI bot that can be indistinguishable from any other human player. Thus it would get much easier to find a 4th player to a game and spare the frustration from somebody leaving one.

Baloot royale is a card game very close to the famous belote game. Some of the main differences in the rules with the classic game is that it doesn’t have an “All trumps” contract. It also has another pass in bidding and one optional “public” card.  Full description of game rules you can find here:

Expected outcome

Ideally we would like to have a smart enough bot which is indistinguishable from a human which can join in anytime when a human is missing. The bot should join a starting game after the bidding has passed and play the game afterwards. Additionally when a player drops in the middle of the game (or is disconnected) then the game should go on smoothly while the bot takes over as a substitute player. The basic role of the bot is to play the game.

Important point is that we don’t need a “super-human” performance. It’s OK for the bot to make mistakes if they are consistent with the behaviour of a regular player.  The important challenge is not to outperform humans but to make moves which are not “annoyingly irrational” from a regular play perspective. Said otherwise we would like the bot to be as close as possible to passing the Turing test – to be indistinguishable from a human player.

The idea is not to come up with a rule-base engine but to build a model that can play as good as a good player. It should also be able to jump in an existing game straight away and finish it in as human-like form as possible (i.e. without many mistakes) – though it should not be that good as to be annoying! Optionally, it should be possible to control the strength of the bot so it can “mimic” noob, medium and advanced players.



Research Problem

The game can be categorized in the set of Imperfect Information Games that is finite, with perfect recall and zero-sum. An article on possible part of the technique to tackle the subject matter can be found here. You can learn more about the actual game from here.


Data description

Imperia Online has provided us with historical games data of real players. Data will be provided with granularity – each played hand in a game. In addition a player ranking is also available.

It contains logs of games of good and bad players’ moves at different hands of game and at different contracts in a sorto of categorical time series.


*player rank (for example good/bad)

*contract (all trumps, no trumps, spades, hearths, diamonds, clubs)

*contract level (normal, double, redouble)

*side of contract (contract taker, partner of contract taker, contract breaker)

*hand number (1,2,3,4,5,6,7,8)

*hand opener (yes/no)

*card type played (A, K, Q, J, 10, 9, 8, 7)

*card color played (spades, hearths, diamonds, clubs)

*result of game (won contract, lost contract, contract break success, contract break fail)

You can download  sample data here…

And also here…

You can download full data set here:
Rooms Logs

Rooms Finished


Share this

Leave a Reply