Occasional Exception Thrown running 2 AIs

[codefactor]

I'm new to this library - but I found a possible issue - and it doesn't always happen due to randomness. I am basically running a game over and over again with AI on both sides until the user asks to join. I've noticed that there is an error during the state.move()

Here is some javascript - if you let it run for a little while (5 minutes or so) you will get the issue near the end of the game. I am guessing perhaps it was supposed to be a draw.

Here is the simple test code:

Live Example: http://jsfiddle.net/MB9kk/

(function(){
    var g, flags;
    function start() {
        flags = 1;
        g = p4_new_game();
        next();
    }
    function next() {
        if ((P4_MOVE_FLAG_DRAW | P4_MOVE_FLAG_MATE) & flags) {
            console.log(P4_MOVE_FLAG_MATE & flags ? 'Mate!' : 'Draw!');
            start();
        } else {
            var mv = g.findmove(5);
            flags = g.move.apply(g, mv).flags; // The move function throws an error
            setTimeout(next, 10);
        }
    }
    start()
})();

The error output is like this:

Uncaught TypeError: Cannot read property '91' of undefined 

And the line is at engine.js:389

function p4_parse(state, colour, ep, score) {
    var board = state.board;
    var s, e;    //start and end position
    var E, a;       //E=piece at end place, a= piece moving
    var i, j;
    var other_colour = 1 - colour;
    var dir = (10 - 20 * colour); //dir= 10 for white, -10 for black
    var movelist = [];
    var captures = [];
    var weight;
    var pieces = state.pieces[colour];
    var castle_flags = (state.castles >> (colour * 2)) & 3;
    var values = state.values[other_colour];
    var all_weights = state.weights;
    for (j = pieces.length - 1; j >= 0; j--){
        s = pieces[j][1]; // board position
        a = board[s]; //piece number
        var weight_lut = all_weights[a];
        weight = score - weight_lut[s];  // <-------- This line is throwing the error

Temporarily what I've done is to put a try { } catch around the call to .move() and if there is an exception, I make it so the current player has resigned.

Note - I am only using the engine part with my own entirely different display.

[douglasbagnall]

Thanks codefactor. This is a good report.

Uncaught TypeError: Cannot read property '91' of undefined

Is it always 91? I think that is a rook's corner, which possibly points to problems with the castle checks.

If you want to remove the randomness, you could search for a problematic game a bit like this:

(function(){
    var g, flags;
    var seed = 1;
    function start() {
        console.log("using random seed " + seed);
        flags = 1;
        g = p4_initialise_state();
        p4_random_seed(g, seed);
        p4_fen2state(P4_INITIAL_BOARD, g);
        seed++;
        next();
    }
  //...

Then when you get the error, if you change the initial seed to the logged number, it should occur right away.

I don't have time to look at it right away, but I will try to soon.

[codefactor]

Thank you for replying so quickly - The number 91 is not consistent from one run to the next.

In the code example its hard to follow the game from just the logs - but on my real page you can see the game and the 2 AI players duke it out until there are no pieces left on one side except the black player's king. I can't remember all of the other pieces left on the white side.

[codefactor]

Using your way of setting the random seed will give the error on the 2nd time - so setting the seed to 2 will do the trick. Here is the fiddle: http://jsfiddle.net/MB9kk/1/

[douglasbagnall]

Aha. Does this version work better?
https://github.com/douglasbagnall/p4wn/blob/b99a9e0ca037f794f258bf2e85422a0c99a117b9/src/engine.js

That's from before commit 660dcac, which tries to recognise a draw from insufficient material.

Edit: the raw JS is here: https://raw.github.com/douglasbagnall/p4wn/b99a9e0ca037f794f258bf2e85422a0c99a117b9/src/engine.js

[codefactor]

Seems still to give the same problem. I pasted the code from the link and ran it in jsFiddle, same error appears to happen.

[douglasbagnall]

It looks like these are queening moves. Firefox and Chromium play different games (which is a bug in itself).

With the random seed 2, as per http://jsfiddle.net/MB9kk/1/:

• Chrome ends with a white queen promotion to checkmate (63 f8=Q#).
• Firefox comes to a black queen promotion (48 ... a1=Q), which isn't the end of the game. The default display code continues past this without reporting any error.

[douglasbagnall]

This is the problem:

            var mv = g.findmove(5);
            flags = g.move.apply(g, mv).flags;

findmove() returns [start, end, score], while move() wants start, end, promotion, where promotion is the desired promotion (i.e. queen, knight, whatever) if the move gets a pawn to the end. So this use of apply is wrong, but that only shows up when you're trying to promote a pawn. Unless the score returned by findmove() happens to represent a valid piece, move() will try to use an undefined lookup table.

I can see the attraction of symmetry between findmove and move, but the score can actually be useful to findmove() callers (so they can, e.g., decide to try again with a deeper search). At the same time, it seems to
me that findmove ought to return the chosen promotion, rather than relying on the display side to assume a queen (which is the only thing findmove considers, but that could change one day).

So the upshot is this works (http://jsfiddle.net/MB9kk/2/) :

            var mv = g.findmove(5);
            flags = g.move(mv[0], mv[1], P4_QUEEN).flags;

and I won't change the signatures of the functions before the next big version, so I'll close this.