Logo how does it work ?


Castle AI utilities a weighted argument based approach as described in this video.
A Castle AI Bot will take into consideration a number of arguments (made by mustachioed arguers!) when making a pick.


Quality
The predetermined value of a card based on CubeTutor Global Pick Orders. Black Lotus = 100, Pale Moon = 1
Strategy
As the draft progresses, the Bot will gain a better understanding of what strategy it is drafting to and will pick cards in accordance with that strategy. For example Mono Green Ramp or Dimir Control.
Signal
A measure of how "open" each colour is, based on the quality of the cards that have been passed to the bot.
Synergy
A measure of how well a given card works with those cards that the bot has already drafted. For example if the bot has previously picked Voltaic Key then Time Vault will score well for synergy.
Colour
The strength of commitment a card has to a particular colour. Cryptic Command = Very Strong, Wurmcoil Engine = None.

Arguer Importance

As the draft progresses, the relative importance of each Arguer changes. For example Signal is very important Pack 1, Pick 2 but slowly diminishes over the course of the first half of the pack until it is no longer relevant. The graph below shows how Arguer Importance changes over the course of a typical 3 pack, 15 pick draft.

Rationalising a pick



For each card in a pack, each relevant arguer will analyse it and provide a strength score to a governing party - The King, who will rationalise all of the arguments into a pick.

You can interrogate the decision process for any given pick on the Draft Bot Reasoning page when you have a draft running. To access that page, simply click on one of the Castle drafters on the draft table plan.

You will see a table like this.

You can see that for each card in a pack, each Arguer has made an argument. The number inside square brackets shows the importance of the Arguer for this particular pick. A more important Arguer will have more say in the final decision. The colour of the light shows you how strong their argument is. Bright green being a Very Strong argument, and bright red being Very Weak.

A light with no colour means that the Arguer does not favour this card at all.

In the above example Liliana of the Veil was chosen, because she fits our Strategy, has Synergy with our previous picks, is a high Quality card and because black has a decent Signal.

Scenarios

Every now and then you will see an extra item of reasoning pop up in the table.

This indicates some sort of special circumstance, or additional information that the bot will consider when reviewing this card.

So far we have the following scenarios

  • Draft Around Me - The card poses some attributed that can lead the bot into forcing a particular direction, for example Storm or Reanimator.
  • Bomb - Reserved for cards in the top 5%. These are picked by Castle AI (almost) every time.
  • On Colour Bomb - Also only considers cards in the Top 5%. This will be used late on in the draft when our colours are established.
  • Hate Draft Bomb - Also only considers cards in the Top 5%. Applies some weight to hate drafting off colour bombs.
  • Combo - Highlights a combo with a card that we've already drafted.
  • On Colour Combo - Highlights a combo with a card that we've already drafted, and the target card is On Colour.
  • Off Colour Combo - Highlights a combo with a card that we've already drafted, and the target card is Off Colour.

Reporting a pick

There are a lot of cards and strategies in Magic: The Gathering! Castle AI has been written in such a way that it's principles can be generally applied to most situations, but more importantly that it can be easily extended. If you feel that a pick violates general principles of good drafting then you can report it. Some things to consider before doing so:

  1. There is often not a "correct" pick, particularly early on in a draft when decisions are often informed by personal bias. Please avoid reporting picks based on your own personal preferences. P1,P1 Time Warp over Time Walk is clearly a questionable decision, however Goblin Guide over Garruk Wildspeaker isn't necessarily incorrect just because you prefer the ramp strategy.
  2. It isn't possible to include every niche strategy in Castle AI. If your cube supports "Mono Blue Goblin Enchantress", well it's possible that Castle AI may not pick that up. We're certainly interested in hearing about such strategies though!
Thumb Up     Thumb Down

To see picks that have been reported by others, head over to the Pick Decision Reports page. Anyone can vote on whether a pick report makes sense or not, and we will work on tackling the "most voted for" issues.