sgf-viewer: Update unhardened games + add baseline attack games

[tomtseng]

• Replace games against cp505-v1 with new games run with better configs
• Replace A-MCTS vs. cp505-v64 games with A-MCTS-R vs. cp505-v8 games — with better configs, our adversary from the unhardened training run no longer performs well against >= 16 visits
  • A-MCTS-R gets a much better win rate against cp505-v8 than A-MCTS-S, and with only 8 victim visits, A-MCTS-R is still computationally tractable
• Add some baseline attack games
  • Edge attack vs cp505-v8 (some wins when attacking as white)
  • Edge attack vs cp505h-v1 (no longer wins)
  • Mirror attack vs. cp505-v1 (occasionally wins)

[netlify[bot]]

✅ Deploy Preview for goattack ready!

Name	Link
🔨 Latest commit	ea8128c785ef2512949157a7a3a2d0b25a3eb1c0
🔍 Latest deploy log	https://app.netlify.com/sites/goattack/deploys/6392b63041e7100008689cef
😎 Deploy Preview	https://deploy-preview-25--goattack.netlify.app/
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify site settings.

[kellinpelrine]

Something may be breaking in Game Analysis -> How the victim's predicted win rate varies over time. The moves look really strange and it won't let me go past move 143. Picture below:

[kellinpelrine]

Game Analysis -> Summary is currently empty. Do you have something intended to be here? If not I can write something.

I'd also consider adding a summary for the baseline attacks, since as we keep adding more content to the website could become harder for audience to see big picture without help. E.g. (maybe there's better way to phrase this though):

In this section we examine simple, no-learning attacks. These test the robustness of KataGo to some types of unsophisticated but likely out-of-distribution play. We find these attacks are generally ineffective against the hardened version of KataGo, although the mirror go attack still gets some wins at low visits. Overall, to find consistent weaknesses, a more powerful approach like ours seems necessary.

[ed1d1a8d]

@kellinpelrine I think having a summary for the Game Analysis would be helpful, feel free to take a stab at writing one. And I like your summary for the baseline section. Will add to the site.

[ed1d1a8d]

Fixed https://github.com/HumanCompatibleAI/KataGoVisualizer/pull/25#issuecomment-1343725939 via 621b96c.