4 ms·

I'm sorry for your bad experience. "Undo and redo" shortcut don't seem to work, but the clickable undo and redo button at the top left work. It's a quick proj

I'm sorry for your bad experience. "Undo and redo" shortcut don't seem to work, but the clickable undo and redo button at the top left work. It's a quick project which arose from curiosity : "what would happen if the hateful AI tried to do the opposite". This AI does its best to always provide you some outs to the limit of its abilities. You have plenty of freedom to guide the game into interesting landscape while having an AI that always offer you safe ways to recover. But you still have to play your part, the AI can only handle half the job. Lovetris has some interesting complexities, and touch the subject of collaborative AI whereas Hatetris is an adversarial AI. Collaborative AI and AI alignment are interestingly hard research subjects which have philosophical and technological components. With subject as deep as love and working jointly, it's hard to satisfy everybody. Lovetris with its "exhaustive short horizon exploring strategy with simple terminal heuristics" isn't capable of perceiving long-term intent farther than its tree searching depth. And reciprocally, if you set the horizon too far, the human player is bad at reading the intent of the AI, and doesn't do what the AI suggest. There are probably better strategies, like monte-carlo tree search with some neural networks which could better solve the intent communication problem between the human and the AI, and provide a better experience, but they take a little more work than just flipping a sign. Add a little feedback via data collection and such an AI can probably send its player into an hedonistic loop.

raldi 5y ago
I think all that's needed here is a note on the website clarifying that you're getting the best piece to complete a line right now, and that this will frustrate anyone looking to set up for a tetris.
- GistNoesis 5y ago
  Thanks, I added some notes to not disappoint new players.
ShamelessC 5y ago
All due respect - I wouldn't really consider this to be an example of anything close to approaching the subject of AI alignment - at least not in terms that people talk about in the past few years. "Exhaustive short horizon exploring strategy with simple terminal heuristics" - is this in reference to the for loop which calculates the future possible states of the game up to a certain depth? It seems to be a state exploration algorithm. The number of possible game states is well-defined here. Perhaps you're conflating the subject of video game AI with concepts related to machine learning AI?
- GistNoesis 5y ago
  The AI is optimizing for a different goal than the player : Like a loving helicopter mom, this AI is trying its best to help you reduce tower height now (to reduce risk) whereas the human player has different objective often related to increasing its score over a certain horizon. The thing is from the programming point of view there are not a unique way to define what should be a "best move".
  - ShamelessC 5y ago
    Are you responding to someone else? My issue is not with understanding the algorithm's flaws and the difficulty surrounding providing players with an optimal choice considering players have different goals than the heuristics used in the algorithm. My issue is with you using incorrect terms and conflating machine learning with a relatively simple graph based algorithm with a relatively simple heuristic. There is no machine learning going on in this codebase whatsoever as far as I can tell, but you seem to be claiming there is. Do you care to respond to that?
    - GistNoesis 5y ago
      >There is no machine learning going on in this codebase whatsoever as far as I can tell, but you seem to be claiming there is. Do you care to respond to that? Afaik I haven't coined the term "machine learning" in this thread and made no such claim. To clarify : There is no machine learning going on in this codebase (though the problem is interesting enough that you can replace the simple AI by a more advanced one). I was just merely pointing that AI alignment problem already affect even simple decision making agents.

raldi 5y ago

I think all that's needed here is a note on the website clarifying that you're getting the best piece to complete a line right now, and that this will frustrate anyone looking to set up for a tetris.

GistNoesis 5y ago
Thanks, I added some notes to not disappoint new players.

ShamelessC 5y ago

All due respect - I wouldn't really consider this to be an example of anything close to approaching the subject of AI alignment - at least not in terms that people talk about in the past few years. "Exhaustive short horizon exploring strategy with simple terminal heuristics" - is this in reference to the for loop which calculates the future possible states of the game up to a certain depth? It seems to be a state exploration algorithm. The number of possible game states is well-defined here. Perhaps you're conflating the subject of video game AI with concepts related to machine learning AI?

GistNoesis 5y ago
The AI is optimizing for a different goal than the player : Like a loving helicopter mom, this AI is trying its best to help you reduce tower height now (to reduce risk) whereas the human player has different objective often related to increasing its score over a certain horizon. The thing is from the programming point of view there are not a unique way to define what should be a "best move".
- ShamelessC 5y ago
  Are you responding to someone else? My issue is not with understanding the algorithm's flaws and the difficulty surrounding providing players with an optimal choice considering players have different goals than the heuristics used in the algorithm. My issue is with you using incorrect terms and conflating machine learning with a relatively simple graph based algorithm with a relatively simple heuristic. There is no machine learning going on in this codebase whatsoever as far as I can tell, but you seem to be claiming there is. Do you care to respond to that?
  - GistNoesis 5y ago
    >There is no machine learning going on in this codebase whatsoever as far as I can tell, but you seem to be claiming there is. Do you care to respond to that? Afaik I haven't coined the term "machine learning" in this thread and made no such claim. To clarify : There is no machine learning going on in this codebase (though the problem is interesting enough that you can replace the simple AI by a more advanced one). I was just merely pointing that AI alignment problem already affect even simple decision making agents.