M2M Day 368: What are my options?

This post is part of Month to Master, a 12-month accelerated learning project. For October, my goal is to defeat world champion Magnus Carlsen at a game of chess.

Yesterday, I ran into a bit of a problem: My chess program, as it currently exists, won’t be able to help me become a grandmaster. Instead, it will only help me correctly identify chess moves as “good” or “bad” 70% of the time, which is worse than I can do as an amateur human chess player.

So, if I want to have any shot of defeating Magnus, I need to try something else. In particular, I can either continue searching for an effective algorithm or I can try to find a completely new approach.

Before trying something new, I think it’s certainly worth continuing down the algorithmic path. After all, the failure of this one particular algorithm doesn’t imply that the entire general approach is hopeless.

In fact, I would have been extraordinarily lucky (or clever…?) if the first algorithm (or class of algorithms) I tried was a functional one. Most often, when trying to create new technical solutions, it takes a few iterations, or more, to land at a workable solution.

Of course, these aren’t random iterations. I still need to be smart about what I try next…

So, what should I try next?

Within the algorithmic space, there are only two possible paths I can go down, and they aren’t exactly mutually exclusive:

1. Build a better model

Use the same dataset, but build a more robust model using Convolutional Neural Nets. Start with a highly sophisticated model, even if it’s not reasonable for human execution, just to see if a high-performance algorithm can even be made. If it can, then slowly reduce the sophistication of the model until it performs just at the level of Magnus Carlsen, and hope that this (or some optimized version of this) is learnable and executable by a human. If it can’t, then it means that my dataset is crappy, and I need to try option #2

2. Build a better dataset

Create a new dataset, constructed and labelled based on some new principle. In particular, try to use the fact that I’m human to my advantage. Maybe, construct the dataset based on a set of heuristics (i.e. for each piece: the number of other pieces directly attacking it, the number of other pieces able to attack it within one move, the number of possible moves it can make, the relative value of the piece compared to the other pieces on the board, etc.). Develop a long list of possible heuristics, and then use the computer to help find which ones are useful and in what combination (this is effectively what the convolutional neural net, from #1, should do, but it will likely be impossible to unravel what heuristics it’s actually using). Alternatively, use the same kind of board input, but label the outputs differently. Perhaps, for example, only try to output which piece is the best to move in each position.

Here’s my plan for moving forward:

1. Construct a Convolutional Neural Net, and run my current dataset through it.
2. In this way, either validate or invalidate the goodness of my dataset.
3. Assuming (very strongly so) that the problem is with my dataset, construct a new dataset, and run it through the Convolutional Neural Net.
4. Continue this process until I find a dataset that is good.
5. Then, start stripping back the model (i.e. the Convolutional Neural Net) while trying to maintain the performance.
6. If the model can be reduced, so that it can be performed by a human, without overly compromising the performance, ride off into the sunset (and by that I mean… spend a ridiculous amount of time and brainpower learning and practicing the resulting algorithm).
7. If the model cannot be reduced without sacrificing performance, try to find another good dataset.

Honestly, this whole process is quite thrilling. In a different life, maybe I would turn this into a Ph.D. thesis, where I try to algorithmically solve chess and then generalize these findings to other domains (i.e. How to develop computer-created, human-executable decision-making algorithms).

But, for now, the Ph.D. will have to wait, and instead, I’ll see what I can accomplish in an hour per day over the next few days.

Read the next post. Read the previous post.

Max Deutsch is an obsessive learner, product builder, guinea pig for Month to Master, and founder at Openmind.

If you want to follow along with Max’s year-long accelerated learning project, make sure to follow this Medium account.