Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Now this is a more of a computer science problem than a math problem but it is concerning game theory and it does seem a lot like nim (it's from an online judge), so i'm kinda stuck on this one i'd really appreciate any help you could give me.

You have $n$ piles of stones, the $i$-th pile has $a_i$ stones on it; you also have a list of available moves, where moves are described by "remove $x_1$ stones from pile $1$, $x_2$ stones from pile $2$, $x_3$ stones from pile $3$, etc., etc." ($x_i$ can be $0$). This is a two-player game where each player in his/her turn can choose one move and play it, and the player that can't make any move loses. For a given configuration you should determine if the winner is the first or the second player.

share|improve this question
    
Assuming both players play optimally of course. –  Dragan Markovic Feb 4 '13 at 19:05

1 Answer 1

You're correct that this is a nim variant. More generally, all impartial games (those where both players have the same moves available to them) of this sort fall under the broad category of the Sprage-Grundy theorem which says that their values are equivalent to the value of some nim-pile.

But for determining who's to win from a given 'single-position' configuration, you don't need any of the broader information; instead, you just need the concepts of 'next player wins' ($\mathcal{N}$-positions) and 'previous player has won' ($\mathcal{P}$-positions), along with the rules for determining the value of a position from all of its options: if any of the moves available from a given position is to a $\mathcal{P}$-position, then that position is an $\mathcal{N}$-position (since the next player can win by moving to the $\mathcal{P}$-position). Otherwise (i.e., if all of the positions that can be reached from the given position are $\mathcal{N}$-positions), the position is a $\mathcal{P}$-position; there are no good moves available.

This evaluation can then be used in one of two ways; if you're interested in the value of a single position then you can simply build the tree (more accurately, DAG — Directed Acyclic Graph) of moves from that position (being careful to cache 'transpositions' - e.g., since making move $m_1$ followed by $m_2$ leads to the same position $P_{12}$ as making move $m_2$ followed by $m_1$, you don't want to compute the value of $P_{12}$ twice) and evaluate it.

If you're going to be computing the value of many positions then it will generally be better to build a table of values, but in games which have many piles this is likely to be a very large table. Since all of the moves have to decrease one or more coordinates, the table can be built outwards from the origin in 'shells': first fill in the value for the origin (which by definition will be a $\mathcal{P}$-position, since there's no legal move to make). Next, fill in the values for all cells with the sum of their pile-sizes equal to 1 (i.e., $(1,0,0,\ldots),\ (0, 1, 0, \ldots),\ (0, 0, 1,\ldots),\ \ldots$); since any move decreases one or more coordinates it has to decrease the sum of the pile-sizes and so you'll already have processed all of the positions a given position's value can depend on before you process that position itself. Note that for a fixed number of dimensions (i.e., piles), it's easy to write the iteration over each shell as a series of nested loops with appropriate limits. For instance, the four-pile case would look like:

for ( totalSum = 0; totalSum <= 16; totalSum++ ) {
  for ( pile0 = 0; pile0 <= totalSum; pile0++ ) {
    for ( pile1 = 0; pile1 <= totalSum-pile0; pile1++ ) {
      for ( pile2 = 0; pile2 <= totalSum-(pile0+pile1); pile2++ ) {
        pile3 = totalSum-(pile0+pile1+pile2);
        ...
      }
    }
  }
}

If the number of piles is given as part of the input you'll have a bit more work to do, but there are still relatively straightforward ways of iterating over all combinations summing to a given number.

share|improve this answer
    
The size of the table will be about $\prod_{i=1}^n (1+a_i)$ unless there are modular constraints to exploit (say the only move for a pile with $73$ is to remove $10$ stones, then that's equivalent to a pile of size $7$ with a $1$-stone move). The DAG will also be about that size if the move sets are arithmetically dense. –  Erick Wong Feb 4 '13 at 19:20
    
Thanks for the fast answer, anyway i forgot to mention that there is a constraint that says that the sum of sizes of all piles is less than 16. So i guess the table is the way to go.I know about the DAG method, but how do you build said table?I would upvote you but it requires 15 rep :( –  Dragan Markovic Feb 4 '13 at 19:26
    
@DraganMarkovic I've updated my answer with a description of how to build such a table. –  Steven Stadnicki Feb 4 '13 at 19:58

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.