This example illustrates how to use TensorFlow.js to perform simple reinforcement learning (RL). Specifically, it showcases an implementation of the policy-gradient method in TensorFlow.js with a combination of the Layers and gradients API. This implementation is used to solve the classic cart-pole control problem, which was originally proposed in:
- Barto, Sutton, and Anderson, "Neuronlike Adaptive Elements That Can Solve Difficult Learning Control Problems," IEEE Trans. Syst., Man, Cybern., Vol. SMC-13, pp. 834--846, Sept.--Oct. 1983
- Sutton, "Temporal Aspects of Credit Assignment in Reinforcement Learning", Ph.D. Dissertation, Department of Computer and Information Science, University of Massachusetts, Amherst, 1984.
It later became one of OpenAI's gym environmnets: https://github.com/openai/gym/blob/master/gym/envs/classic_control/cartpole.py
The gist of the RL algorithm in this example (see index.js) is:
- Define a policy network to make decisions on leftward vs. rightward force given the observed state of the system. The decision is not completely deterministic. Instead, it is a probability that is converted to the actual action by drawing random samples from binomial probability distribution.
- For each "game", calculate reward values in such a way that longer-lasting games are assigned positive reward values, while shorter-lasting ones are assigned negative reward values.
- Calculate the gradients of the policy network's weights with respect to the actual actions and scale the gradients with the reward values from step 2. The scale gradients are added to the policy network's weights, the effect of which is to make the policy network more likely to select actions that lead to the longer-lasting games given the same system states.
For a more detailed overview of policy gradient methods, see: http://www.scholarpedia.org/article/Policy_gradient_methods
For a more graphical illustration of the cart-pole problem, see: http://gym.openai.com/envs/CartPole-v1/
- Allows user to specify the architecture of the policy network, in particular, the number of the neural networks's layers and their sizes (# of units).
- Allows training of the policy network in the browser, optionally with simultaneous visualization of the cart-pole system.
- Allows testing in the browser, with visualization.
- Allows saving the policy network to the browser's IndexedDB. The saved policy network can later be loaded back for testing and/or further training.
yarn && yarn watch