Support "Human-Like" model

[TristanCacqueray]

We need to pass a new -human-model parameter after the KataGo command, and it looks like such wrapper script works:

#!/bin/sh
command=$(echo $1)
shift
exec /nix/store/g398d059f5g8s7i6k5grrx7zl21cf4pk-katago-1.15.3/bin/katago $command -human-model /nix/store/6dhgn7sxxvy78s9bd50kzazz3n1dxhfm-b18c384nbt-humanv0.bin.gz $*

Is there any plan to support this natively in katrain?

Thanks for the tool by the way!

[simonguoxm]

For me, it is not necessary to support this.

[pdeblanc]

I've gotten this working locally. Here's what I did:

• Download the b18c384nbt-humanv0.bin.gz model into my ~/.katrain directory
• Create a copy of katrain/KataGo/analysis_config.cfg called katrain/KataGo/humanlike_config.cfg
• In humanlike_config.cfg, append the line humanSLProfile = preaz_2d (i.e. emulating 2-dan human play from the old days; other settings are possible)
• In katrain/core/engine.py, around line 120, add the following lines:

             cfg = find_package_resource(config["config"])
+            if "human" in model:
+                cfg = cfg.replace("analysis", "humanlike")
             exe = self.get_engine_path(config.get("katago", "").strip())

It is necessary to have two separate configs, because the default KataGo models will raise an error if the humanSLProfile setting is present in the config.

I could consider putting up a PR, but I'm not sure if this solution is sufficiently user-friendly:

• You have to switch the model in the General/Engine settings to get humanlike play (e.g. with the Policy Weighted AI)
• This will also use the human policy for analysis; there's no convenient way to use the human policy for play while simultaneously using the AI policy for analysis
• There should probably be a way to change the humanSLProfile via a menu instead of editing the config. This could include the following options and suboptions:
  • Modern amateur style
    • select rank
  • Pre-AI amateur style
    • select rank
    • select opponent rank (e.g. to see how a human 4d might play against a 10k opponent)
  • Professional style
    • select year (1800 to 2023)

For reference, the humanSLProfile options are documented here: https://github.com/lightvector/KataGo/blob/master/docs/Analysis_Engine.md

[PJTraill]

way to change the humanSLProfile

The opponent rank is also supported for emulating post AlphaZero amateurs.

documented here

They are also summarised in the supplied sample configuration: https://github.com/lightvector/KataGo/blob/v1.15.0/cpp/configs/gtp_human5k_example.cfg

[pdeblanc]

OK, I'm sort of thinking about putting together a PR for this feature, but where to cram the new options? We need a drop-down menu (to select Modern Amateur, Pre-AI Amateur, or Professional style), and up to two numeric input fields, depending on which style is selected.

[PJTraill]

while simultaneously using the AI policy for analysis

You say there is no convenient way to do this, but surely the -human-model commandline argument does that by default? That is what https://github.com/lightvector/KataGo/blob/master/docs/Analysis_Engine.md#human-sl-analysis-guide says.

[pdeblanc]

while simultaneously using the AI policy for analysis

You say there is no convenient way to do this, but surely the -human-model commandline argument does that by default? That is what https://github.com/lightvector/KataGo/blob/master/docs/Analysis_Engine.md#human-sl-analysis-guide says.

Katrain's AI is not a simple wrapper around KataGo. There are various AI modes which use KataGo's analysis in different ways to select a move (cf. https://github.com/sanderland/katrain/blob/master/katrain/core/ai.py).