Monday 16 Nov 2020
==================

bouldering
----------

Morning bouldering session from 07:15 to 09:00.
Met up with Pete and his friend James.
I warmed up on blues/greens and practiced on the yellows.

We met Tom. Tom is a bouldering legend! He showed us how to do some of the more challenging climbs, which seemed impossible to us mere mortals.

Things to try next time:

- Try the grip strength wall at the start
- End with the pull-up bar
- Can I make the yellow dyno?
- Try the yellow cave climb without feet.
- Try the green cave climb like Tom did it.
- Keep practicing on the yellows, try an orange or two.

open source
-----------

I received an email from Alexandr. He wants to maintain stayradiated/termcolors. I haven't touched that project in years. I added him as a maintainer -- I'm  interested to see what he does.

machine learning
----------------

Instead of working on Runn I spent the day learning about MuZero - a model for machines to learn to play games. 

I started out without much idea of how MuZero worked. What I hoped for was that I could train it to play board games really well.

I initially started out by trying to implement the rules for the Arboretum board game [[ https://www.boardgamegeek.com/boardgame/140934/arboretum ]], however I realised it would be better to start with a simple game. I wrote a simple Tic-Tac-Toe game in python, which was a fun learning experience in itself (I haven't written much Python).

I found this excellent series of articles on how MuZero works: How To Build Your Own MuZero AI Using Python [[ https://medium.com/applied-data-science/how-to-build-your-own-muzero-in-python-f77d5718061a ]]

After getting a rough understanding of how MuZero worked, I knew that it was going to be difficult challenge to implement it myself. Luckily, I found the [[ https://github.com/werner-duvaud/muzero-general ]] project. This is a ready to use implementation of MuZero, complete with several games already configured -- inculding a Tic-Tac-Toe game.

I ran the code locally but my laptop doesn't have a supported GPU so the training was very slow.

I tried a variety of other services but didn't have much luck.

- google colab:
  - free
  - very easy to get started
  - can on CPU, GPU or TPU
  - seemed to freeze and crash if left alone for too long
  - difficulty interacting with muzero-general CLI
- gradient
  - I couldn't figure out how to just launch a server and run some code
- aws
  - lots of server types to choose from
  - wouldn't let me launch a server, I had to get permisson from AWS to increase my quota
- datacrunch.io
  - simple to use
  - cheap
  - prepaid credit
  
I'm currently training MuZero on Tic-Tac-Toe on the cheapest DataCrunch server. I'll leave it overnight and see how well it did tomorrow.

media
-----

[[ https://fast.ai ]] looks a good resource to learn more about nerual networks.

This is a good explanation of the Monte Carlo Tree Search [[ https://www.youtube.com/watch?v=UXW2yZndl7U ]].

lunch
-----

Went out for lunch with Ivan & Tim. We ordered food from SteamPunk Labs and ate back in the office. Talking about how overly complicated web development is these days.

exercise
--------

Around 16:00 I went for a run up & down the car park building in bare feet.
Listening to the Exponent podcast with Ben & James [[ http://exponent.fm ]].

dinner
------

- Naked Kitchen Thai Jungle Curry
- Rice crackers & hummus
- Hot cocoa

todos
-----

- [x] Send invoice to Mish Guru.

journaling
----------

Is quite different on a laptop. I get distracted quite easily. I may need to switch off the internet next time...