Timeline for Is it unethical to make a game AI that is secretly non-competitive?
Current License: CC BY-SA 3.0
17 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Oct 25, 2017 at 20:52 | comment | added | Casey | @LordFarquaad Yes, I meant more that you can't really do it to the computer so much (since generally it involves throwing out some unsafe option based on the behavior you expect of your opponent) with its instantaneous reaction times. | |
| Oct 25, 2017 at 20:50 | comment | added | Lord Farquaad | @Casey try telling an Amiibo that. If you repeat the same behavior in the same situation often enough, you most certainly will face a "hard read". After all, as the AI gets more confident that you'll respond with a certain option, the risk of leaving the other options open drops pretty quickly. | |
| Oct 25, 2017 at 19:56 | comment | added | Casey | @NicolBolas Not exactly... I mean the concept of a "hard read" or a "mixup" doesn't really apply to the computer. | |
| Oct 25, 2017 at 19:46 | comment | added | Nicol Bolas | @Casey: In the fighting game example, that's essentially how high-level play works: bait out and punish anything unsafe, while only doing things that are safe. It's simply that, for the computer, what constitutes "safe" is a lot more narrow than for a human. | |
| Oct 25, 2017 at 19:07 | comment | added | Casey | To your point, I feel like, even if the AI is beatable doing "perfect" play, it ends up requiring you to play in an unnatural way tailored to beating it (like in your fighting game example to only ever wait for the AI to attack and try and stuff its attacks or punish it for throwing something unsafe out). | |
| Oct 22, 2017 at 6:26 | comment | added | Tobia Tesan | @Ray In fact, MTCS is particularly relevant in this discussion because it is an anytime algorithm, and its performance can be easily tweaked by increasing or reducing its computational budget (i.e. time or nodes expanded). I have zero experience with doing something like this in practice, but I don't think it would be conceptually hard to increase or decrease its budget depending on how well it's faring, in order to be a match to the player's ability at all times. | |
| Oct 21, 2017 at 19:22 | comment | added | Ray | @Carl That'd be AlphaGo, one of the reasons I needed an "until recently" in my comment. An earlier version of it beat the number 2 player in the world last year. | |
| Oct 20, 2017 at 16:47 | comment | added | Carl | @Ray off topic, but at about the time of your comment, google released the fact that they made a Go AI that taught itself so well, that not only can it destroy every other AI, but it plays the game in ways we have never conceived. | |
| Oct 20, 2017 at 15:37 | comment | added | Lord Farquaad | Very much this. As an anecdote, my friends and I were playing an RTS (Rise of Nations) against a few computers with no trouble. Then my friend modded it so the population cap went from 200 to 2000, and the same computer slaughtered us, even when it was 3 of us vs one computer. We humans had diminishing returns on controlling more units at a time, while the computer didn't. | |
| Oct 20, 2017 at 11:42 | comment | added | Dulkan | Strategy game AIs are much worse than "perfect as far as i know" though. Yes, they have perfect micro, which turns them into a significant challenge for most players, but they are also very rleiant on their script and don't respond well to unorthodox strategies. E.g. in Warcraft 3, you could defeat the hardest AI with a single worker, becuase the AI cannot handle a tower rush at all. It ignores the worker building the tower until it is too late, then proceeds to waste its units vs those towers, instead of taking the enemy base. | |
| Oct 19, 2017 at 22:33 | comment | added | Ray | @TemporalWolf Every possible move in a 7 or fewer piece endgame is still quite a few orders of magnitude away from "every possible move". | |
| Oct 19, 2017 at 21:48 | comment | added | TemporalWolf | @Ray That may have been true ten years ago, but it's less so now: we have pre-computed solution tables for all 7 piece and less endgames. Once there are fewer than 8 pieces there is a distinct best move somewhere in those 170TB tables. | |
| Oct 19, 2017 at 21:46 | comment | added | ttbek | That's why the groups that were working on Go are now turning to SCII. ^_^ | |
| Oct 19, 2017 at 17:23 | history | edited | Nicol Bolas | CC BY-SA 3.0 | added 189 characters in body |
| Oct 19, 2017 at 17:22 | comment | added | Nicol Bolas | @Ray: "If the AI can play perfectly and still lose, then a human who plays perfectly may also lose. That's not ideal either." "Perfect" here being "within the boundaries of the game in question". I gave RTS's as an example because the genre often uses Fog of War, which prevents you from knowing everything. And without perfect knowledge, you cannot have genuinely perfect play. You can only have "perfect" play within the boundaries of what you know. | |
| Oct 19, 2017 at 17:19 | comment | added | Ray | If the AI can play perfectly and still lose, then a human who plays perfectly may also lose. That's not ideal either. Rather, you want a game that has sufficient complexity and/or hidden information that determining a perfect strategy is impossible or impractical. This is why humans still beat computers at Chess and Go until recently; the computer can't analyze every possible move fully, and must rely on heuristics, much like the human relies on intuition. | |
| Oct 19, 2017 at 14:56 | history | answered | Nicol Bolas | CC BY-SA 3.0 |