5 ms·

I wonder if AI self-play has any implications for e-sports balance. Could balancing be aided, or at least improved, by ultra-strong AI players, reducing the cha

I wonder if AI self-play has any implications for e-sports balance. Could balancing be aided, or at least improved, by ultra-strong AI players, reducing the chance of broken mechanics making it to live?

jdoliner 8y ago
Sure seems like it would. A big part of DOTA balance these days is Icefrog (DOTA's BDFL) talking to pro players and about the meta-game and observing pro play to figure out what should change. Having pro-level AIs that you could just set lose to play 100,000 games and then come back to you with statistics about the current patch would probably help a lot. I think it would also raise the level of play that we see from pros. That's what happened with chess bots when they got good enough to be the best human players.
- shawn 8y ago
  One specific thing that bots could help Icefrog solve: Many games are draft wins, i.e. a >90% chance of winning solely due to the heroes you've picked. There's almost no point in playing these games out for 20-40 minutes; it's boring for everyone involved. This will always exist in dota, but being able to simulate hundreds of years of drafting + games could help reduce this.
  - Benjammer 8y ago
    They're already made great strides into expanding the pro-level meta hero pool over the last several major patches. As an observer it feels like the viable pro hero pool is bigger than ever this year at TI8.
  - trocadero 8y ago
    >Many games are draft wins, i.e. a >90% chance of winning solely due to the heroes you've picked. Can you explain?
    - danielvf 8y ago
      Each team of five picks the hero’s they will play the game with. This is like an extra complicated game of rock-paper-scissors. At the end of the picking, one teams composition may be be so setup to exploit the weakness of the other, that the actual 30-40 minute game is almost pointless.
      - trocadero 8y ago
        You pick back and forth though, right? How does one screw up the draft so badly as to have only 10% chance of winning?
        shawn 8y ago
        There are so many drafting combinations that it's not quite obvious it's a bad idea until you hit the 3 minute mark. Take OpenAI game 3 as an example. The first two games, OpenAI wiped the floor with the humans and taunted them that they had a >90% chance of winning. The third game, OpenAI was saying the bots had an >80% chance of losing by 5min. The sole difference was the draft.
        bigger_cheese 8y ago
        Not mentioned was each team gets to ban out 5 heroes as well. This is a bigger factor in my opinion. Each team alternates two bans, then two picks, then there is another round of two alternating bans followed by two picks and finally a single ban and pick round for the 5th hero. Various in-meta heroes are usually "first pick/ban worthy" which means they tend to get picked or banned in the first phase and tend to shape the rest of the draft as teams will build the core of their strategy around the first phase heroes or around countering oppositions first phase heroes. Another strategy is to avoid "showing your hand" during first phase by first picking strong but generic heroes that can fit into many potential lineups to keep opponent guessing. This leads to a lot of mind games where even commentators don't know what role the hero is going to be played in until the culminating 5th pick when the draft comes together. Some teams are very good at specific strategies or have certain players exceptionally skilled at individual heroes which necessitate certain first phase bans against them lest they have an advantage. For instance If a team is known for having a player good at the hero "Wisp" it will often force out a first phase Wisp ban from opponents because it is the kind of hero that when played well can be absolute nightmare to play against. In some ways I find the draft mini game to be just as interesting as the main game especially in the longer tournaments where you can see new metagames emerging as captains adjust their pick strategies.
  - prestonh 8y ago
    I really doubt that at TI level the drafting advantage is significant enough to give one team 90% advantage. My guess would be closer to 2/3, granted thats still just a complete guess. I think often analysts will use the draft as an excuse to in retrospect justify the outcome of the game, but that doesn't bear on the reality of the situation at all. Though, I would be really interested to see some data that shows the significance of drafting.
minimaxir 8y ago
Training prices would have to drop significantly. It's much cheaper to take a PR hit and apply a hotfix instead of spending $100k+ on building a self-play AI.
- adw 8y ago
  If it were really just 100k it would be a huge win to have the AI. That's one developer for three to six months. (I'd handicap the crossover somewhere between one and two orders of magnitude more expensive than that.)

jdoliner 8y ago

Sure seems like it would. A big part of DOTA balance these days is Icefrog (DOTA's BDFL) talking to pro players and about the meta-game and observing pro play to figure out what should change. Having pro-level AIs that you could just set lose to play 100,000 games and then come back to you with statistics about the current patch would probably help a lot. I think it would also raise the level of play that we see from pros. That's what happened with chess bots when they got good enough to be the best human players.

shawn 8y ago
One specific thing that bots could help Icefrog solve: Many games are draft wins, i.e. a >90% chance of winning solely due to the heroes you've picked. There's almost no point in playing these games out for 20-40 minutes; it's boring for everyone involved. This will always exist in dota, but being able to simulate hundreds of years of drafting + games could help reduce this.
- Benjammer 8y ago
  They're already made great strides into expanding the pro-level meta hero pool over the last several major patches. As an observer it feels like the viable pro hero pool is bigger than ever this year at TI8.
- trocadero 8y ago
  >Many games are draft wins, i.e. a >90% chance of winning solely due to the heroes you've picked. Can you explain?
  - danielvf 8y ago
    Each team of five picks the hero’s they will play the game with. This is like an extra complicated game of rock-paper-scissors. At the end of the picking, one teams composition may be be so setup to exploit the weakness of the other, that the actual 30-40 minute game is almost pointless.
    - trocadero 8y ago
      You pick back and forth though, right? How does one screw up the draft so badly as to have only 10% chance of winning?
      - shawn 8y ago
        There are so many drafting combinations that it's not quite obvious it's a bad idea until you hit the 3 minute mark. Take OpenAI game 3 as an example. The first two games, OpenAI wiped the floor with the humans and taunted them that they had a >90% chance of winning. The third game, OpenAI was saying the bots had an >80% chance of losing by 5min. The sole difference was the draft.
      - bigger_cheese 8y ago
        Not mentioned was each team gets to ban out 5 heroes as well. This is a bigger factor in my opinion. Each team alternates two bans, then two picks, then there is another round of two alternating bans followed by two picks and finally a single ban and pick round for the 5th hero. Various in-meta heroes are usually "first pick/ban worthy" which means they tend to get picked or banned in the first phase and tend to shape the rest of the draft as teams will build the core of their strategy around the first phase heroes or around countering oppositions first phase heroes. Another strategy is to avoid "showing your hand" during first phase by first picking strong but generic heroes that can fit into many potential lineups to keep opponent guessing. This leads to a lot of mind games where even commentators don't know what role the hero is going to be played in until the culminating 5th pick when the draft comes together. Some teams are very good at specific strategies or have certain players exceptionally skilled at individual heroes which necessitate certain first phase bans against them lest they have an advantage. For instance If a team is known for having a player good at the hero "Wisp" it will often force out a first phase Wisp ban from opponents because it is the kind of hero that when played well can be absolute nightmare to play against. In some ways I find the draft mini game to be just as interesting as the main game especially in the longer tournaments where you can see new metagames emerging as captains adjust their pick strategies.
- prestonh 8y ago
  I really doubt that at TI level the drafting advantage is significant enough to give one team 90% advantage. My guess would be closer to 2/3, granted thats still just a complete guess. I think often analysts will use the draft as an excuse to in retrospect justify the outcome of the game, but that doesn't bear on the reality of the situation at all. Though, I would be really interested to see some data that shows the significance of drafting.

minimaxir 8y ago

Training prices would have to drop significantly. It's much cheaper to take a PR hit and apply a hotfix instead of spending $100k+ on building a self-play AI.

adw 8y ago
If it were really just 100k it would be a huge win to have the AI. That's one developer for three to six months. (I'd handicap the crossover somewhere between one and two orders of magnitude more expensive than that.)