Getting used to playing against HaChu 0.19

The human results are getting better. White_horse is beating it most of the time. Kenjin has beaten it a couple of times.

H.G.Muller (Test_DontPlayMe), the author of HaChu, observed that it doesn’t see the danger of piece pile-ups when it is pinned. Inspired by this, last night I decided to play it on evens, despite not yet having beaten it with 2 Kings.

I played straight-forwardly, without a special strategy. As soon as it brought it’s Lion out into the centre, a fight broke out. I concentrated on pinning its pieces against the lion and ganging up on them. This way I acquired a small material advantage. The centre fight continued and my advantage increased until I got a second Lion. Then the programmed stupidity kicked in during the endgame, and I had an easy finish.

Programmed stupidity

On 81Dojo today kenjin was playing HaChu 0.19. The following position arose when he was short of time:

 

Computer finds a mate - human doesn't

Computer finds a mate – human doesn’t

The computer spotted mate, and so started sacrificing everything it could to delay it. But kenjin didn’t actually find the mating sequence, and took a long time to win, despite hachu having sacrificed lion, dragon horse and white horse.

 

Given time, can you find the mate (mate in the sense that there is no defence to checkmate, not in the mating-problem sense where every move is check)?

The basic castle, and a slightly stronger one

Usually in Chu Shogi, the King remains in the centre, and the surrounding step movers are used to form a castle. This is typically composed of the Drunk Elephant, the two Blind Tigers, and one or two Gold Generals. Here is a typical setup, which I call the basic castle:

Showing the weakness of the basic castle to a lion and vertically moving piece.

Showing the weakness of the basic castle to a lion and vertically moving piece.

Note that if the central defender is pinned to the King by a vertically-moving piece (here a White Horse), then a Lion can give checkmate by sitting on the head of either Blind Tiger (shown here on 8j, but 6j is also possible). For this reason, a side-moving piece is needed on the third rank.

Although it takes an extra move to set up, it is often preferable to swap the position of the Drunk Elephant and the right-hand Blind Tiger. We can see why here:

Showing only half the weaknesses remain.

Showing only half the weaknesses remain.

The same mate is possible, but this time only on 8j. The Drunk Elephant protects 6j. So I call this the enhanced castle.

HaChu 0.19 is proving to be strong

The results against strong humans were very impressive today. And there are a few more weaknesses in the endgame to be patched too.

These changes came about, because I have been proposing to fit the evaluator from my ancient Chu 2.11 program to HaChu. This evaluator meant that Chu 2.11 played really beautiful opening play (in comparison to HaChu 0.18, or Zillions of Games – not compared to a human!). But it’s search was quite weak. So the prospect of combing my evaluator with HaChu’s search engine looked interesting.

So this week HGM began explaining to me how HaChu worked, and I showed him my evaluator. We began discussing my piece tables (which were carefully designed to enable reasonable opening play without the step movers getting in the way of the ranging pieces, and vice versa.). From these discussions, HGM tried some of the principles out in HaChu 0.19, and the effect was surprisingly (to me) large.

With further relatively easy improvements to come, I wonder how much (if any) of an improvement implementing my piece tables will make. The only way to find out is to try, and currently I’m planning to have a go after Easter.

One interesting point about these changes – HGM tells me that HaChu 0.19 loses to HaChu 0.18.

Heroic defeat

Last night, I had the privilege of seeing four kings on the board.  Blind_tiger, a really strong player, was playing HaChu at 60-10 (this is a small time handicap for the human player, as currently the bot can’t use byo yomi).  The bot managed to get a second lion, and blind_tiger was faced with the unenviable prospect of playing 10 seconds a move when 17 points down (81dojo’s assessment).

 

He played brilliantly, and went over 10 points ahead. But both Drunk Elephants promoted to Crown Prince, and he lost his way in the (I presume) unfamiliar situation. Still, it was a heroic defeat.

Here is the game record.