Hacker News Reader

Comment by nomilk | Hacker News Reader

AlphaZero (Google's chess bot) famously started with no 'book knowledge' (e.g. ideas on how to play chess), instead, it learned by playing against itself many, many times over.

Has anyone done something similar to that with Monopoly? (by contrast, the technique used in this video is much more basic; it comes up with some specific rules a player will use, then has those bots play eachother)

View on HN

I have thought it couple of times, most recently after this video. I so want to make it since I feel all these videos tend to lack two things the zero-style would offer.

One: it would offer no human assumptions.

Two: more importantly it should and could include the trading part. (the op's video had some, but it felt really really limited from what I gathered)

It is really rare to see trading to be implemented. Yeah, it is the hard part but in competetive monopoly, it is big part.

I am yet to figure both input/output spaces for it. As soon as I figure them, I probably will do proper zero-style ai for many games.

Isn't trading always blocked? Whenever you would allow someone to build a complete Color it's better not to allow it, and whenever you also would it's only advantageous if your color is stronger than the other thus the other party should never agree.

Any other exchange becomes pointless as you won't be completing a color anyways.

You seem to reason in terms of trying at all costs to block the other players, not to maximize your gain. I agree the ultimate goal is to do better than all the others, but the tactics to get there may involve temporarily favoring other players in win-win exchanges to grow quicker, or to block a third player already far ahead, etc.

In the second case, the stronger color can be offset by giving the weaker color more cash / other properties.

There is also the case when you don't kind a want to buy the property and it goes to action. Now you can bid again, hoping for cheaper prices or to take risk and try to drive to price as high as possible for the actual buyer. Same possibility exists for any auction.

This was tested in the video. Players that don't trade lose. Refusing trades where the opponent gets a better set underperforms always trading to complete your sets.

As I have let this simmer in my mind, one really great way to drop the interactions needed to complete auction, could be ducth action. I think it is called that. (the auction where everone submits their maximum that they are willing to bet in secret. Winner of the action is highest bet, and they pay the second highest bet as winning bet.)

This could reduce the input space and define it such that it allows this thing. Hmm.

Reply on news.ycombinator.com