Re: [gnugo-devel] Strong non-linear behavior?

[SP LEE]

> Thus, the black played excellent to save his group successfully, yet played 
> tenuki at the last moment!
> I also saw such behavior earlier in other games.
> 
> The moves to win the semeai were valued very high in the beginning stage, 
> and then suddenly dropped to less than the real territorial value, about 
> half of the beginning value. To my understanding, even at the final victory 
> of the semeai, there is whole lot of strategy values in it. For instance, 
> the out side opponent groups will be all weakened.
> 
> Anyone has some suggestion as which part of the program should I check?
> 
> SP Lee
> 

I have just traced the gnugo's reasoning for playing tenuki at the end of the 
semeai and the following code in the function connection_value made me 
completely confused.

    float old_burden = 2.0 * crude_weakness_a * soft_cap(sizea, 15.0);

    /* The new burden is the burden of defending new joint dragon; but
     * we share this burden proportionally with the other dragon.
     */
    float new_burden = 2.0 * crude_weakness_sum * soft_cap(sizea + sizeb, 15.0)
         * sizea / (sizea + sizeb);

    return_value = 1.05 * (old_burden - new_burden);
    /* If ahead, give extra bonus to connections. */
    if (margin > 0.0)
      return_value *= 1.0 + 0.02 * margin;

The situation is that dragon a is alive and dragon b is undergoing semeai, 
which must be connected to dragon a. Now it looks like gnugo found the "new 
burden" is bigger than "old burden" and didn't value this connection. Does it 
mean that crude_weakness_sum is calculated wrong?

SP Lee