I am newbie on AI and I am just thinking about the possibility of adding "reinforce learning" algorithms to WL. It would be a perfect companion to the actual genetic algorithms and it would help to improve our strategies already optimized and backtested, but "learning" from incoming out-of-sample data.
¿Is it suitable?
There are several codes in the public domain that can help improve the incorporation of Reinforcement Learning (RL). Example, OpenAI has a reinforcement learning platform called "Spinning Up" that provides source code for reinforcement learning algorithms. There are also open source libraries like TensorFlow and PyTorch that can be used to implement RL algorithms.
There are various reinforcement learning techniques, such as Deep Q-Networks (DQN), Actor-Critic, Q-Learning, SARSA, and more, and there are various open source resources to enable developers to build and train reinforcement learning agents in C#.
¿Is it suitable?
There are several codes in the public domain that can help improve the incorporation of Reinforcement Learning (RL). Example, OpenAI has a reinforcement learning platform called "Spinning Up" that provides source code for reinforcement learning algorithms. There are also open source libraries like TensorFlow and PyTorch that can be used to implement RL algorithms.
There are various reinforcement learning techniques, such as Deep Q-Networks (DQN), Actor-Critic, Q-Learning, SARSA, and more, and there are various open source resources to enable developers to build and train reinforcement learning agents in C#.
Rename
Admitting I know less than anyone about the subject.. conceptually, can you explain how "reinforced learning" is any different than what the Evolver already does?
I don't see such a major architectural change as in the cards for the Evolver. But could be the basis for another extension in the future.
It would be as an extension of the Evolver, or a complement, the difference is that instead of genetical algorithms this would use a learning method based on rewards and penalties according to the results while interacting with the environment.
In plain words: it could be used to optimize existent strategies with fresh out of sample data. Our strategies are optimized and backtested using data from the past, despite we take a large portion out of sample it could be overfitted or the environment can change. Using a reinforcement approach on an existing strategy could be useful to correct deviations adding/supressing rules according to actual results or modify variables to adapt to changes in the market.
In plain words: it could be used to optimize existent strategies with fresh out of sample data. Our strategies are optimized and backtested using data from the past, despite we take a large portion out of sample it could be overfitted or the environment can change. Using a reinforcement approach on an existing strategy could be useful to correct deviations adding/supressing rules according to actual results or modify variables to adapt to changes in the market.
If you start using “out of sample” data for your criteria, it suddenly isn’t “out of sample” data any more.
Indeed, and there is still risk of overfitting, but it's a different approach to optimize a strategy and with the possibility of making it adaptative to changes of market regime.
Your Response
Post
Edit Post
Login is required