1800 my @ss

[fern]

To date I have only seen one program more or less responding to what you expect of such or such elo......
Thje rest are a mix of GM moves and moron moves.. Maybe someone believes an average player is got adding genuses with idiots.

My best
Fern

[DomLeste]

OK whos better with limited Elo feature Delfi or Hiarcs?

[Dr.Wael Deeb]

OK whos better with limited Elo feature Delfi or Hiarcs?

To get Fernando straight to the point,eh

[cpuchess]

That seems a little ridiculous. Seems like the program shouldn't be making blunders like that for anything above 1500 or so...

[fern]

Dunno about Delfi, but Sjeng make the trick well....

Fern

[fern]

In my very humble opinion, you get relatively decent results if you simply shorten the deep of search. It is the trick Sjeng uses and it works well. It is specially good when the player to simulats is a 1300-1500 elo. 2-3 ply search is more or less how a player of such a force play.

My best
fern

[Tony Thomas]

Dunno about Delfi, but Sjeng make the trick well....

Fern

I like the trick used by Sjeng, leaving a piece stranded and allowing me to take it while I am way behind. Then there are times it defeats me in 20 moves.

Trying to get past Sjeng 1400 level regards
Tony

[M ANSARI]

I think the best way to play a computer is to limit it in ply. I have had a lot of fun with Rybka 2.3.2a at ply 1 and 2 and 3. After that it becomes extremely tough to beat. It is hard to make an engine emulate a human player's strength ... engines will always be very strong tactically ... while humans generally play quite good positionally. I am not sure how the ELO feature works in Rybka ... but I would choose ply 1 as a start and see how that goes. At ply 1 I would guess it plays about 1800 ELO level is you are playing also very quickly.

[bob]

In my very humble opinion, you get relatively decent results if you simply shorten the deep of search. It is the trick Sjeng uses and it works well. It is specially good when the player to simulats is a 1300-1500 elo. 2-3 ply search is more or less how a player of such a force play.

My best
fern

It only works for a narrow range of Elo expectations, and it is not uniform in how the reduced depth affects Elo. I tried this on my cluster last year and gave up. A 3 ply search will see some tactics a 1400 player will miss, but it will also miss some things, particularly in an endgame, than an 800 player would not miss.

The other _major_ issue with just reducing the depth is that you are still using an evaluation that is over 2000. To produce a 1500 level player, using a 2000 level eval, you have to play 1000 level tactics. That is not how a real 1500 player "feels' when you play against one.

I have an option in crafty that works pretty well, although the scale is not linearly proportional to Elo. But it has been tested with a couple of hundred thousand games to verify that against computers, I can reliably drop the elo by 200, or by 400, or whatever. Only problem is, I can reliably drop it, but I can't drop it to a specific Elo rating because the "drop" is only proportional to the Elo of the program at full strength on that specific hardware...

The nice thing about the current approach I use is that the program doesn't feel quite as artificial. It begins to miss more tactically, make an occasional obvious mistake, etc, just like a real 1400 player will.

[JVMerlino]

With Chessmaster we took a completely different approach. Certainly less researched and scientific, and not sure if the results are more or less satisfying to a human opponent.

We simply made several dozen different personalities and then played as many games as possible with them against each other and against human USCF rated players.

We definitely didn't play enough games to be reasonable accurate. Most personalities are probably within 100 points USCF, but history now shows that some play several hundred points above or below their ratings. We simply played the games, did the math and assigned a rating to each personality -- no verification was done of the final numbers because we didn't have time.

Also, as I'm sure Dr. Hyatt will agree to, the farther away we got from optimal settings, the wider the variance on the personality's rating.

jm