Which is better, strategy devised by Machine Learning or Human Intuition?


#1

Hi all, I am very new to Halite III (and machine learning). I would like to hear from the community, which is the best way to devise the strategy?

Basically I am trying to get some sense, if using a black box method through machine learning, or using a white box method through human intuition and hard coding the strategy, is the better way?

I can see the white box method is straightforward and easy to tell the intention behind. But I am not sure if the machine learning black box way has an inherent advantage over the human intuition white box way?

Thanks for sharing your thought with me!


#2

You can hit platinum with just a good movement, some regular math algorithm for ship targets finding and manually handling some situations. However I’m not sure what is going on at diamond rating. Their bots is just crazy compare to everything else. But if you are a newbie in such challenges it would be near impossible to pass top 40-60 border anyway.


#3

I’m a beginner at machine learning so take this with a grain of salt, but here’s my view:

When the machine learning model is devised by some of the world’s top ML experts with vast resources at their disposal, then for some games it seems there is an advantage in training the model from first principles with no interference from human intuition. AlphaGo Zero versus AlphaGo Master provides a clear example.

Whether this is true for all games is not clear to me. It’s also worth noting that the AlphaGo programs were much stronger than their machine learning components. The ML part was used to guide a Monte Carlo-like search, whereas in this game there probably isn’t enough time per move for anything like that.

With my level of understanding of machine learning, I think I’ll get the best results by ignoring it completely and coding the whole solution based on my own understanding. If there’s a place for any sort of ML in my final entry, it will be for tuning parameters within that human-devised strategy.


#4

If you just check a few moves for each ship, and which one gives you the most halite, that will get you quite far. And that is the start of machine learning, a value (Loss function) to optimize (maximize).

I see you just started on the leaderboard, you can get to #999 easily just by fixing how your ships move around and collect halite. Bogdandm is #50 so he might have some clever ideas on more advanced things like building new dropoffs and adjusting strategy to the board layouts.

The “AI” way would be to make a (deep?) neural network to learn and copy the moves of the top 10 players, but then you’d still have the problem that you don’t know what features to use I think? I think its different from differentiating static dog & cat photos and traditional deep learning tasks.