Jul 6, 2021
A drop in motivation
My plan last week was to make the challenge a round 5 weeks: spend one week on Connect 4, and then, regardless of the results, use the last week to give my best try at Photosynthesis (even if it fails).
I ended up getting bogged down in various tasks:
- I tested modifications to the Connect 4 training process, without much improvement (or so I thought). I was also frustrated by the limitations imposed by the very slow speed.
- I worked on a big update to improve speed by running NN model evaluations in parallel, but the work is still in progress.
- At the beginning of this week, I thought I would take a break from all that, and worked on the game logic for Photosynthesis.
I was affected by my perceived lack of progress (and maybe a slight pressure as the challenge was nearing its end), and lost my focus and motivation. The same thing happened on Day 2 of the challenge: I got overwhelmed and had to take a step back to regain focus.
Regaining focus for the end of the challenge
In order to do that I made the following decision today:
I am temporarily pausing Photosynthesis in order to focus 100% on Connect 4.
This will give me a clear path for the end of the challenge, and let me finish speed and training optimisations on Connect 4 without spreading myself too thin. I am definitely not done with Photosynthesis, though. I already wrote the game logic, and I'm looking forward to come back to it in a future challenge!
Now the game plan for the last few days is:
- Finish the speed optimisation update I started
- Look at the metrics of the previous training runs, come up with improvements to the training process, and test them
- Run one last long training session, and test the results (vs. different basic AIs and human players)
I will run the last session this week-end, and write an end-of-challenge Post on Monday with the final results. Until then, I will focus on the task at hand!
And to finish on a positive note, here is the story of an interesting experience:
My Lee Sedol moment
Lee Sedol is the world champion that first got beaten by AlphaGo in 2016. This week-end I had my own (very scaled-down) experience of that moment, when I first got beaten by my Connect 4 AI. Of course, I promptly played a rematch and won ;)
Although I am still playing better than it, it managed to get a few wins from me (as well as from another friend who played it). It didn't fall for any cheap trick, and playing it felt like playing a human.
This showed me that I had made more progress than I gave myself credit for. And indeed, the new version beats the basic AI with a 9/1 win/loss ratio (up from 3/1 at the beginning of the week).
There is still a big margin for progress of course, but reaching that first milestone was an exciting moment!
