Bellman's Curse on Advice - dust-nib
In dynamic programming and reinforcement learning, the most fundamental challenge we face is just grokking the high dimensionality of a state space. Go is a ...
In dynamic programming and reinforcement learning, the most fundamental challenge we face is just grokking the high dimensionality of a state space. Go is a “simple” game whose rules can be fully explained on a single sheet of paper, yet those rules yield more board configurations than the number of atoms in the universe (~10174). Given the state space, how does any solver for this game even represent these configurations, let alone evaluate them? While games are helpful toys to start with, this phenomenon of “exponential explosion” in systems is extremely common in many real-world problems,…
related reading
- Curse of dimensionality - Wikipediaen.wikipedia.org
- What's going on here, with this human? - Graham Duncan Bloggrahamduncan.blog
- Evolution as Backstop for Reinforcement Learning · Gwern.netgwern.net
- Debugging Reinforcement Learning Systemsandyljones.com
- Elon Litman | Elements of a Vector Spaceelonlit.com
- A Free Lunch from the Noise:Provable and Practical Exploration for Representation Learningarxiv.org
- Thousand-dimensional structure — Resolutionresolution.org
- A Theory of Deep Learning | Elements of a Vector Spaceelonlit.com
- Dynamical Mean Field Theory, Random Matrices and Learning in High Dimensionsyoutube.com
- Embedded Agency (full-text version) — LessWronglesswrong.com
- Bellman equation - Wikipediaen.wikipedia.org
- Deep Reinforcement Learning: Pong from Pixelskarpathy.github.io