Neural networks for simulating play conditions

edited March 2006 in Story Games
So I was reading this thread on the Forge, where people are discussing the merits of mathematical to supplement/partially obviate playtesting. This got me thinking about something I'm reading up on at the moment, neural networks.*

It seems to me that you could set up a structure (e.g., I want 2 inputs to go into every conflict, with one, stat, having a bigger impact than the other, skill), and then plug in a ton of cases with desired outcomes.
That is, feed in a number of conflicts of varying obstacles, and in each case say what you want the likelihood of success to be (between 0 and 1). As I understand it, the network should then be able to adjust the weights it gives to the original inputs to give you the outcome you desire from the conditions you force on it.

This might be enlightening. For example, it might suggest that the common value that the player and their adversary get to add to their conflict score (e.g., the number of dice you add to your stat+skill) needs to be contributing between 30-50% of the players total - any more, and the variability leads to multiple conflicts becoming too risky.

Or I might be blowing smoke from my ass. I wonder if anyone savvier has ever thought to go down this road? Just as a way to tweak values, not as a replacement of the play experience, I hasten to add.

*I'm here referring to what my primer - Callan's The Essence of Neural Networks - calls artifical neural networks - ie, they're not attempting to model biological reality, but get a certain kind of output from certain kinds of input.

Comments

  • I've done stuff like this, although I consider more of a feasibility test that happens long before playtesting. I do this sort of thing when I'm composing the rules.
  • Do you use a neural network framework - ie, changes happen at the level of weights between inputs, and occurs through learning algorithms - or another framework? I'm just trying to work out whether the neural learning thing actually lends itself to these situations, or there is a neater way to do this.
  • I am, for once, not that high tech. I usually just make a spreadsheet or database and dicker with the formulae involved until I like what I get.
  • Wow. First comment on this thread and it's tangential...

    I would argue against neural networks for game design unless you're willing to inextricable tie a computer to your mechanics. It might be interesting for a web-mediated game where resolution was mechanic.

    Reason being: first of all, the key unsolved problem in the field is how many neurons you need to simulate a particular system. But more importantly, if you do get the network to simulate your system, it doesn't teach you anything about the system itself.

    In design terms, you set up a little neural network, and you give it the inputs the rest of your system impose on your resolution mechanic. You experiment with different numbers of neurons, and you eventually get outputs that look pretty much like you want. All the network can tell you is that this linear equation emulates your system.

    To be honest, Josh's suggestion is a really good one. For tweaking the math of a system, a smatter of Probabilty and a grounding in Excel (or OpenOffice.org's Calc) work absolute wonders.

  • OK cool. I'm pretty Excel savvy so that should be the most direct route for me. It would have been a nice goal to develop my network knowledge, but now I know that it's the wrong tool for the job!
Sign In or Register to comment.