Author | Nathan Lambert compile | source of vitamin k | forward Data Science

1. Supervised learning

Modern reinforcement learning is almost entirely focused on deep reinforcement learning. The word “deep” in deep reinforcement learning implies the use of neural networks at the heart of the algorithm. Neural networks perform some high-dimensional approximations during learning. Having said that, the model does not need to have many layers and features, which is a common misconception that implies many layers.

Almost all courses and tutorials assume that you can fine-tune simple neural networks to approximate state values or create final strategies. Historically, these models have been highly sensitive to all the following training parameters: learning rate, batch size, model parameters, data normalization, and so on. There are many problems with RL learning, and the best approach is to learn supervised learning and let AutoML do the work for you.

Start by learning some code, understand everything about it, and then rebuild. Do this a few times and you’ll learn well. Supervised learning is also a skill that will translate into most areas of computer science over the next decade, so stay close.

2. Search methods in artificial intelligence

RL is a search problem in the policy space. Some of the most influential papers in recent years have been heavily inspired by the age-old search problem. Let’s take a look at three of the most influential recent papers at RL:

  1. The most mainstream reinforcement learning result to date: Deepmind mastered many games and conducted massive reinforcement learning. More recently, future actions are being explored using planning networks (deepmind.com/research/pu…)

  2. Current status of model-based RL research: Model-based Policy Optimization (MBPO). MBPO is searching the vicinity of the state action space for a more complete knowledge. This analog knowledge is like a simple exploration. (arxiv.org/abs/1906.08…).

  3. Research status of MODEL-free RL: Soft Actor-Critic (SAC). SAC is known for its combination of effective exploration and peak performance. It does this by maximizing the entropy term on the strategy. As a search problem, direct policy search is the most critical aspect of RL.

You can take ai courses at UC Berkeley and MIT and try a project or two to challenge yourself.

inst.eecs.berkeley.edu/~cs188/

Ocw.mit.edu/courses/ele…

3. Understand academic papers

Note that I didn’t say read academic papers; the key is to be able to understand them.

The academic world is changing rapidly, with papers being published daily in the active Twitter community (seriously, follow a few prominent researchers – it helps to keep you up to date). What you need to learn is to separate influential papers from noise and progress from hype.

It’s not about citations, although I admit that citations are a good way to distinguish. What you should focus on are experiments: do they follow the trajectory of proposing a problem, experimenting, proposing a solution, or are they just incremental improvements on another approach? The best papers seek truth, not numbers.

The best research is to look for the most important questions, regardless of the tide of academic discussion. Learn to recognize the signs, and you’ll be able to keep up with the field and become a valuable asset for the next decade.

At the end

I have written many tutorials on the basic skills of reinforcement learning and will post them later.

  1. What is the Markov decision process?
  2. Linear algebra for reinforcement learning.
  3. A basic iterative approach to reinforcement learning.

Original link: towardsdatascience.com/3-skills-to…

Welcome to panchuangai blog: panchuang.net/

Sklearn123.com/

Welcome to docs.panchuang.net/