Whish list for Rocinante.

[Antonio Torrecillas]

Hi,
I'm starting a new development effort for Rocinante and I want to continue the customer driven development that starts with the release 2.

Rocinante is an also-rans engine which use uncommon algorithms:
- Probability B Star search (PB*)
- Monte Carlo tree search alpha beta (mcts_ab)

PB* is more tactical than mcts_ab, but mcts_ab is stronger (+100 elo). It's strength range is around 1650-1750 elo or weaker depending of settings, hardware and time control.

Those of you that have used it, how it behave ?

feedback script:
Hardware / Settings you're using.
Use: engine-engine human-engine others ?
subjective assessment: opening, mid game, end game. Tactics, strategic play...
How it compares to other engines of similar strength?
How it compares to a top engine dumbed down to a similar strength?
what do you wish for the next release?

if you don't feel good to explain this in public, you can send me a PM with your feeling. (languages: English, French and Spanish are welcome)

There is no wrong answer.
It's feedback and wish time!

Many thanks.
Best regards, Antonio.

[Marek Soszynski]

Hi Antonio,

I would use your engine in infinite analysis because of its different algorithms. Linux compiles would interest me the most.

[Vinvin]

I didn't even know Rocinante is so "special". I'll try, as human, to play some blitz games against it this evening .

[IGarcia]

Hardware: Intel Core2 Duo / Linux 64bits
Play: human vs engine

subjective assessment:
(based in a few games.. about 10)

In early stages sometimes play unexpected moves: for example g6 but the never plays Bb7, instead plays Be7?

After development I find it likes too much to castle longside.
The games I manage to win or draw was same pattern: those ones long castle and later moving pieces into attack leaving the king vulnerable.


-How it compares to other engines of similar strength?
I usually do not play aganst engines, so hard to say.


-How it compares to a top engine dumbed down to a similar strength?
I done some few short matches against stockfish 2.3.1 at level 8, 7, 6
Seems like level 6 or 7 is about equal. Time control was 2'+3"


-what do you wish for the next release?
Keep improving it. Very good work.

-other details:

It plays too fast, even having plenty of time. Also if you set a position and "go infinite" it replays very fast with bestmove, instead looking for better move.

The impression is it plays at aproximated even peace, like a fixed amount of time per move. Unless you tactical blunder, there it plays instantly... letting you know you are in trouble

Thanks for all the efforts.

[IGarcia]

I also played some engine-vs-engine against pulsar-2007-9b-Linux

Only few games, seems about equal, but the particular thing was games repeated, exactly!

Code: Select all

[Event "k1-mc1"]
[Site "?"]
[Date "2012.11.25"]
[Round "?"]
[White "Pulsar2007-9b-Linux"]
[Black "Rocinante 2.0"]
[Result "1/2-1/2"]
[ECO "C47s"]
[TimeControl "2/3"]

1.e4 e5 2.Nf3 Nc6 3.Nc3 Nf6 4.d4 exd4 5.Nxd4 Bb4 6.Nxc6 bxc6 7.Bg5 h6 
8. Bxf6 Qxf6 9.Qf3 Qxf3 10.gxf3 d5 11.exd5 cxd5 12.Bb5+ Bd7 13.Bxd7+ Kxd7 
14.O-O-O Bxc3 15.bxc3 Kc6 16.Rd4 Rhe8 17.Rhd1 Re5 18.c4 Rd8 19.f4 Rh5 
20.cxd5+ Rdxd5 21.Rc4+ Rc5 22.Rxc5+ Rxc5 23.Rg1 g6 24.Rg3 Rf5 25.Rc3+ Kd7 
26.Rf3 Kd6 27.Kd2 Ke6 28.Re3+ Kd5 29.Rd3+ Ke4 30.Rd7 Rd5+ 31.Rxd5 Kxd5 
32.Kd3 c6 33.c3 Kc5 34.h4 Kd5 35.c4+ Kc5 36.Kc3 a6 37.a3 a5 38.Kd3 Kd6 
39.Kd4 c5+ 40.Ke4 Ke6 41.a4 f5+ 42.Kd3 Kd6 43.Kc3 Kc6 44.Kd3 Kb6 
45.Kc3 Ka7 46.Kd3  Ka8 47.Kc3 Kb8 48.Kd3 Kc7 49.Kc3 Kd6 50.Kd3 Ke6 
51.Kc3 Kd7 52.Kd3 Ke7 53.Kc3 Ke8 54.Kd3 Kf8 55.Kc3 Kg7 56.Kd3 Kg8 
57.Kc3 Kf8 58.Kd3 Ke8 59. Kc3 Ke7 60.Kd3 Kd7 61.Kc3 Kc7 62.Kd3 Kc8 
63.Kc3 Kb7 64.Kd3 Ka6 65.Kc3 Kb6 66.Kd3 Kc6 67.Kc3 h5 68.Kd3 Kb6 
69.Kc3 Ka7 70.Kd3 Kb7 71.Kc3 Kb6 72.Kd3 Kc6 73.Kc3 Kd6 74.Kd3 Ke6 
75.Kc3 Kd7 76.Kd3 Kd6 77.Kc3 Ke7 78.Kd3 Kf7 79.Kc3 Kf8 80.Kd3 Kg8 
81.Kc3 Kh8 82.Kd3 Kg7 83.Kc3 Kf7 84.Kd3 Ke8 85.Kc3 Kd8 86.Kd3 Kc7 
87.Kc3 Kb6  {3 fold repetition.}
1/2-1/2

There you can see also some problem in endgame (both engines missed) careless moving the king around to the point Rocinante goes to far away, where a pawn break its possible. Look at 46...Ka8? and 64...Ka6?

I think its possible to replay this game against Rocinante (memorizing it or reading it) and later wait for one of those king moves.

This shows about how fixed is playing Rocinante. Maybe something to look into.

[Antonio Torrecillas]

thanks for the usefull info.
Please, could you repeat your test game against the other algorithm ?.
There is a setting called mcts_ab, to control which of the two is used.
I'm particularly interested on how it plays against human, and how it behave with both algorithms.
If it play too quickly, you can try to increase the parameter probedepth to 2 or 3.

So far that's what I'm scheduling for R2.1
I'm in the process of allowing the engine to support the analysis mode (uci commands:go infinite/stop).
R2 has limited size for the tree. So this is another area to improve.
Solve bug in multiprocessor code in linux.
I'm working in solving tactics, this got me a drop in ELO (-425), but it's coming back slowly (now at -275)
For the evaluation, I've tried several different approaches, and what works well with mcts_ab and/or alpha-beta
doesn't work well for PB*, may be I still need to learn how to construct a true evaluation.
I'll focus on a single algorithm, an evaluation that works well with it, and so on.
In short, I'm pondering the convenience to split the engine in two differents engines.

Best regards,Antonio.