The most human-like engine?

[Odd Gunnar Malin]

I had to run to a webinar with GM Ankit Rajpara so I forgot to mention that the Maia engine with the setup on their site and on Lichess is probably too weak for you. I'm 1700+ and have no problem with it. I guess it is possible to do some adjustment to get it stronger since the engine is lz0. I guess 1900 is lichess rating so 1500+ in real rating.

I played just one game against 1900 Maia at the time control of 10+5 or so and won rather easily, and I am pretty bad at those fast time controls. If they made something like 2100 Maia it might be a good sparring partner for me.

I read somewhere that they had problem to get it play at a higher level, not enough games for learning or something.

[dkappe]

I played just one game against 1900 Maia at the time control of 10+5 or so and won rather easily, and I am pretty bad at those fast time controls. If they made something like 2100 Maia it might be a good sparring partner for me.

An important note about Maia from their github:

The Maias are not a full chess framework chess engines, they are just brains (weights) and require a body to work. So you need to load them with lc0 and follow the instructions here. Then unlike most other engines you want to disable searching, a nodes limit of 1 is what we use. This looks like go nodes 1 in UCI. Note also, the models are also stronger than the rating they are trained on since they make the average move of a player at that rating.

At 1 node — essentially without search — they are very susceptible to tactical oversights.

In my work with small nets, I did a few things differently than Maia:

1. I played out terminal positions from lichess games. Otherwise you end up with nets that don’t know how to play endgames or up a queen.
2. I blended the score with the eval of a shallow SF10 search and created a policy from that same shallow multipv search. This boosts the strength of the net without influencing the “human” feel too much.
3. I tried different sizes of net, all the way down to 2 blocks and 16 filters.

A big net with 100 nodes may play the same rating as a small net with 5000 nodes, but the style — tactical vs positional— may be very different.

So, if you weren’t happy with Maia on lichess, grab yourself the weights and a recent version of lc0 for cpu, then experiment with different nets and nodes per move. It should be an eye opening experience.

[Fritz 0]

I played just one game against 1900 Maia at the time control of 10+5 or so and won rather easily, and I am pretty bad at those fast time controls. If they made something like 2100 Maia it might be a good sparring partner for me.

An important note about Maia from their github:

The Maias are not a full chess framework chess engines, they are just brains (weights) and require a body to work. So you need to load them with lc0 and follow the instructions here. Then unlike most other engines you want to disable searching, a nodes limit of 1 is what we use. This looks like go nodes 1 in UCI. Note also, the models are also stronger than the rating they are trained on since they make the average move of a player at that rating.

At 1 node — essentially without search — they are very susceptible to tactical oversights.

In my work with small nets, I did a few things differently than Maia:

1. I played out terminal positions from lichess games. Otherwise you end up with nets that don’t know how to play endgames or up a queen.
2. I blended the score with the eval of a shallow SF10 search and created a policy from that same shallow multipv search. This boosts the strength of the net without influencing the “human” feel too much.
3. I tried different sizes of net, all the way down to 2 blocks and 16 filters.

A big net with 100 nodes may play the same rating as a small net with 5000 nodes, but the style — tactical vs positional— may be very different.

So, if you weren’t happy with Maia on lichess, grab yourself the weights and a recent version of lc0 for cpu, then experiment with different nets and nodes per move. It should be an eye opening experience.

Is i5 modern or older CPU?

[Eraserheads]

This perennial question has been asked many times before, and I never get tired reading the answers, as it changes with the times.
I keep dozens of engines, most of which are extinct or discontinued, just because I like how 'humanly' they play.

Here are just some of them
1) Chess System Tal II - This engine justifies the name, plays at around a 2400 clip. Sacrifices pieces a lot, with a leaning towards the aesthetic.
2) Zappa Mexico II Dissident Aggressor Settings - Whoever made these settings deserve an award for creating a real wildman of an engine.
Here are the settings:

[ZappaMexicoII]
UCI_ShowCurrLine=2
Threads=auto
Enable Null Move=2
Enable History Pruning=2
Enable Singular Extensions=2
Enable Mate Extensions=2
Enable Futility Pruning=2
Aggressive Futility Pruning=2
TB Min Depth=4
Null Move R4 Min Depth=100
Singular Margin=75
Singular Margin PV=50
Eval Passed Pawn Scoring=125
Eval Pawn Scoring=135
Eval Minor Scoring=70
Eval Major Scoring=70
Eval Kingsafety Scoring=500
Internet Lag=0
Contempt=20
Hide Fail Highs=2
Print PV Tips=1



3) Thinker 5.3b inert - Think Chess System Tal but at a 2800 strength. The sacrifices Thinker creates are more mature, more subtle and often can break even cold blooded chess engines in the 3000 and below range.

4) Bright 0.4a - Think morozevich. This engine is a maverick, and plays in a refreshing original style. Almost always it steers its games towards unusual positions right out of the opening. Plays at a 2800 clip.

[Frank Quisinsky]

Bright and Thinker ...



Bright from Allard Siemelink ... the little "Spark".
More aggressive in open positions.

Thinker was great!
Which such small sources such a strong and interesting program.
A little sensation what the programmer do here!

I remember on a time I am looking on Thinker and SlowChess only (around the same time the programmer of SlowChess comes also with a sensational engine around the corner). So AnMon and SOS are fighting vs. these two programs 24 hours per day. And Frank was happy!

Zappa often a bit boring in style but often enough with very interesting combinations end of mid-games. 75 Elo stronger the x64 as the w32 version. Very interesting program. The setting is new for me. Hm, I have Zappa on my backup drive.

We had a lot of such interesting material.

Best
Frank

[Ajedrecista]

Hello:

[...]

2) Zappa Mexico II Dissident Aggressor Settings - Whoever made these settings deserve an award for creating a real wildman of an engine.
Here are the settings:

[ZappaMexicoII]
UCI_ShowCurrLine=2
Threads=auto
Enable Null Move=2
Enable History Pruning=2
Enable Singular Extensions=2
Enable Mate Extensions=2
Enable Futility Pruning=2
Aggressive Futility Pruning=2
TB Min Depth=4
Null Move R4 Min Depth=100
Singular Margin=75
Singular Margin PV=50
Eval Passed Pawn Scoring=125
Eval Pawn Scoring=135
Eval Minor Scoring=70
Eval Major Scoring=70
Eval Kingsafety Scoring=500
Internet Lag=0
Contempt=20
Hide Fail Highs=2
Print PV Tips=1

[...]

Those settings were made by an user of this forum back in 2010. Here are three early TalkChess threads about Dissident Agressor:

New Zappa Mexico II setting, dead Stockfish

Zappa Mexico II Dissident Aggressor!

Zappa Dissident Aggressor settings

Regards from Spain.

Ajedrecista.

[JohnW]

You could buy a DGT Centaur chess computer that adjusts to your playing strength. Personally I like playing against the old chess computers of the 1980's-90's like the Mephisto, Fidelity and Novag chess computers. Just feels better knowing they're at full strength instead of playing against software on my PC that I have to try and artificially dumb down to my level.

[Chessqueen]

istolacio ...

Other opinion: Most important are the first gaming phase.
Here Dragon by Komodo 3.1 NN is more interesting as Stockfish.
Can be see very easy with "fast won" games / "fast lost" games.

Today we have the possibility to use more as one engine for analyzes.
Interesting are 4 or 8 engines with different strengths and one core.

This combination of engines for analyzes is better as a chess tutor with one engine only.
The main reason why "Playing styles" of engines are all the time important.

Engines for a chess tutor are engines like to play all playing phases with around the same level without bigger weaknesses.
Here RubiChess or Ethereal are good examples. The styles are very balanced.
On other way is to use the strongest available engine!

Example:
I am using for analyzing openings:

- Dragon by Komodo 3.1 NN
- Slowchess Blitz 2.9 NN
- Ehereal
- Revenge 3.0 NN
- Wasp 6.00 NN
- Velvet 6.00 NN
- RubiChess Aug. NN
- often Junior for speculative combinations!

RubiChess is very strong with unbalanced material with many pieces on board.

Best
Frank

Ethereal is showing a great performance in Chess324, it is holding 3rd place after Dragon 3.1, I suppose it has a small net that does not waste too much time looking for unknown positions that it has not been trained for like LCO and great evaluation https://www.chess.com/computer-chess-championship

[supersharp77]

Sorry if this question has already been asked here, but which is, in your opinion, the most human-like chess engine?

WOW...."Most Human Like Engine"......Interesting question....well here's my list....

1 Maia Nets (LC0) 1700..1800..1900...Etc

2 LC0 v0.16.0 Or LC0 0.17.0 or LC0 v0.19.0

3 Human Engine (fully adjustable)

4 Alex Chess

5 Cleo

6 Brama

7 Crafty 11.19 11.23 Crafty 14.11

8 Deep Brute POS 2.1

9 Deep Junior 5-6-7

10 Fairy Max

11 Faile

12 Geko 0.43

13 Kasparov Chess (Java)

14 Thinker 3.0c

15 Lchess 5.0

16 List 4.61

17 Matheus 1.9

18 Obender 3.2x

19 Olipow 2.2

20 Tiffany Chess

21 Plisk

22 Seahorse 1.0

[Fritz 0]

I tried Hiarcs 14 at the fixed depth of 5 and 6 ply. While it's probably stronger than me at 6 ply even if I play at the classical time control (of course, it moves instantly), my impression is that it doesn't resemble a human player very much. It is certainly less human-like than pre-Dragon Komodo, for example.