Tic-tac-toe with only crosses

CJam (58 56 bytes)

2q~:Xm*{7Yb#W=}:F,Xm*{ee{~0a@*\+}%zS*F},_Wf%:z&Mf*1fb:e>

This is incredibly slow and uses a lot of memory, but that's code-golf for you.

Dissection

2q~:Xm*        e# Read input into X and find Cartesian product {0,1}^X
{7Yb#W=}:F,    e# Filter with a predicate F which rejects arrays with a 111
Xm*            e# Take the Cartesian product possible_rows^X to get possible grids
{              e# Filter out grids with an anti-diagonal 111 by...
  ee{~0a@*\+}% e#   prepending [0]*i to the ith row
  zS*F         e#   transposing, joining on a non-1, and applying F
},
_Wf%:z         e# Copy the filtered arrays and map a 90 degree rotation
&              e# Intersect. The rotation maps horizontal to vertical and
               e# anti-diagonal to diagonal, so this gets down to valid grids
Mf*            e# Flatten each grid
1fb            e# Count its 1s
:e>            e# Select the maximum

The number of valid rows is a tribonacci number and is \$\Theta(a^X)\$ where \$a\$ is the tribonacci constant, \$1.83928675\ldots\$. The number of grids generated in the Cartesian product is \$\Theta(a^{X^2})\$; the second filter will reduce the number somewhat, but the intersection is still probably \$\Theta(a^{X^4})\$.


An "efficient" approach ("merely" \$O(X a^{3X})\$) uses dynamic programming. Ungolfed in Java:

public class A181018 {
    public static void main(String[] args) {
        for (int i = 1; i < 14; i++) {
            System.out.format("%d:\t%d\n", i, calc(i));
        }
    }

    private static int calc(int n) {
        if (n < 0) throw new IllegalArgumentException("n");
        if (n < 3) return n * n;

        // Dynamic programming approach: given two rows, we can enumerate the possible third row.
        // sc[i + rows.length * j] is the greatest score achievable with a board ending in rows[i], rows[j].
        int[] rows = buildRows(n);
        byte[] sc = new byte[rows.length * rows.length];
        for (int j = 0, k = 0; j < rows.length; j++) {
            int qsc = Integer.bitCount(rows[j]);
            for (int i = 0; i < rows.length; i++) sc[k++] = (byte)(qsc + Integer.bitCount(rows[i]));
        }

        int max = 0;
        for (int h = 2; h < n; h++) {
            byte[] nsc = new byte[rows.length * rows.length];
            for (int i = 0; i < rows.length; i++) {
                int p = rows[i];
                for (int j = 0; j < rows.length; j++) {
                    int q = rows[j];
                    // The rows which follow p,q cannot intersect with a certain mask.
                    int mask = (p & q) | ((p << 2) & (q << 1)) | ((p >> 2) & (q >> 1));
                    for (int k = 0; k < rows.length; k++) {
                        int r = rows[k];
                        if ((r & mask) != 0) continue;

                        int pqrsc = (sc[i + rows.length * j] & 0xff) + Integer.bitCount(r);
                        int off = j + rows.length * k;
                        if (pqrsc > nsc[off]) nsc[off] = (byte)pqrsc;
                        if (pqrsc > max) max = pqrsc;
                    }
                }
            }

            sc = nsc;
        }

        return max;
    }

    private static int[] buildRows(int n) {
        // Array length is a tribonacci number.
        int c = 1;
        for (int a = 0, b = 1, i = 0; i < n; i++) c = a + (a = b) + (b = c);

        int[] rows = new int[c];
        int i = 0, j = 1, val;
        while ((val = rows[i]) < (1 << (n - 1))) {
            if (val > 0) rows[j++] = val * 2;
            if ((val & 3) != 3) rows[j++] = val * 2 + 1;
            i++;
        }

        return rows;
    }
}

Pyth, 57 51 49 bytes

L.T.e+*]Ykbbsef!s.AMs.:R3ssmyBdsm_BdCBcTQsD^U2^Q2

Like @PeterTaylor's CJam solution, this is brute-force, so it runs in O(n22n2) time. The online interpreter doesn't finish within a minute for n=4.

Try it here for N<4.

Try the diagonals function.

L.T.e+*]Ykbb         y(b): diagonals of b (with some trailing [])
s e                  sum of the last (with most ones) array such that
f                    filter lambda T:
 ! s .AM                none of the 3 element sublists are all ones               
   s .:R3               all 3 element sublists
   s s                  flatten
   myBd                 add the diagonals
   sm_B d               add the vertically flipped array and transpose
   CBcTQ                array shaped into Q by Q square, and its transpose
 sD ^U2 ^Q2             all binary arrays of length Q^2 sorted by sum

C, 460 456 410 407 362 351 318 bytes

This is a really bad answer. It's an incredibly slow brute force approach. I'm trying to golf it a bit more by combining the for loops.

#define r return
#define d(x,y)b[x]*b[x+y]*b[x+2*(y)]
n,*b;s(i){for(;i<n*(n-2);++i)if(d(i%(n-2)+i/(n-2)*n,1)+d(i,n)+(i%n<n-2&&d(i,n+1)+d(i+2,n-1)))r 1;r 0;}t(x,c,l,f){if(s(0))r 0;b[x]++;if(x==n*n-1)r c+!s(0);l=t(x+1,c+1);b[x]--;f=t(x+1,c);r l>f?l:f;}main(c,v)char**v;{n=atol(v[1]);b=calloc(n*n,4);printf("%d",t(0,0));}

Test Cases

$ ./a.out 1
1$ ./a.out 2
4$ ./a.out 3
6$ ./a.out 4
9$ ./a.out 5
16$

Ungolfed

n,*b; /* board size, board */

s(i) /* Is the board solved? */
{
    for(;i<n*(n-2);++i) /* Iterate through the board */
            if(b[i%(n-2)+i/(n-2)*n]&&b[i%(n-2)+i/(n-2)*n+1]&&b[i%(n-2)+i/(n-2)*n+2] /* Check for horizontal tic-tac-toe */
                    || b[i] && b[i+n] && b[i+2*n] /* Check for vertical tic-tac-toe */
                    || (i%n<n-2
                            && (b[i] &&b [i+n+1] && b[i+2*n+2] /* Check for diagonal tic-tac-toe */
                                    || b[i+2*n] && b[i+n+1] && b[i+2]))) /* Check for reverse diagonal tic-tac-toe */
                    return 1;
    return 0;
}

t(x,c,l,f) /* Try a move at the given index */
{
    if(s(0)) /* If board is solved, this is not a viable path */
            return 0;
    b[x]++;
    if(x==n*n-1) /* If we've reached the last square, return the count */
            return c+!s(0);

    /* Try it with the cross */
    l=t(x+1,c+1);

    /* And try it without */
    b[x]--;
    f=t(x+1,c);

    /* Return the better result of the two */
    return l>f?l:f;
}

main(c,v)
char**v;
{
    n=atol(v[1]); /* Get the board size */
    b=calloc(n*n,4); /* Allocate a board */
    printf("%d",t(0,0)); /* Print the result */
}

Edit: Declare int variables as unused parameters; remove y coordinate, just use index; move variable to parameter list rather than global, fix unnecessary parameters passed to s(); combine for loops, remove unnecessary parentheses; replace && with *, || with +; macro-ify the 3-in-a-row check