MADDPG | Pytorch implementation of the MARL algorithm, MADDPG, which correspondings to the paper "Multi-Agent | Reinforcement Learning library

Vectorized Environments are a method for stacking multiple independent environments into a single environment. Instead of executing and training an agent on 1 environment per step, it allows to train the agent on multiple environments per step.

Usually you also want these environment to have different seeds, in order to gain more diverse experience. This is very useful to speed up training.

I think they are called "vectorized" since each training step the agent observes multiple states (inserted in a vector), outputs multiple actions (one for each environment), which are inserted in a vector, and receives multiple rewards. Hence the "vectorized" term

It is indeed true that sampling is not a differentiable operation per se. However, there exist two (broad) ways to mitigate this - [1] The REINFORCE way and [2] The reparameterization way. Since your example is related to [1], I will stick my answer to REINFORCE.

What REINFORCE does is it entirely gets rid of sampling operation in the computation graph. However, the sampling operation remains outside the graph. So, your statement

.. how does the gradient passes through a random operation ..

isn't correct. It does not pass through any random operation. Let's see your example

Horizon refers to how many steps into the future the agent cares about the reward it can receive, which is a little different from the agent's time to live. In general, you could potentially define any arbitrary horizon you want as the objective. You could define a 10 step horizon, in which the agent makes a decision that will enable it to maximize the reward it will receive in the next 10 time steps. Or we could choose a 100, or 1000, or n step horizon!

Usually, the n-step horizon is defined using n = 1 / (1-gamma). Therefore, 10 step horizon will be achieved using gamma = 0.9, while 100 step horizon can be achieved with gamma = 0.99

Therefore, any value of gamma less than 1 imply that the horizon is finite.

As we talked about in the comments, it seems that the Keras-rl library is no longer supported (the last update in the repository was in 2019), so it's possible that everything is inside Keras now. I take a look at Keras documentation and there are no high-level functions to build a reinforcement learning model, but is possible to use lower-level functions to this.

• Here is an example of how to use Deep Q-Learning with Keras: link

Another solution may be to downgrade to Tensorflow 1.0 as it seems the compatibility problem occurs due to some changes in version 2.0. I didn't test, but maybe the Keras-rl + Tensorflow 1.0 may work.

There is also a branch of Keras-rl to support Tensorflow 2.0, the repository is archived, but there is a chance that it will work for you

It could be a problem with your Python version: k-armed-bandits library was made 4 years ago, when Python 3.9 didn't exist. Besides this, the configuration files in the repo indicates that the Python version is 2.7 (not 3.9).

If you create an environment with Python 2.7 and follow the setup instructions it works correctly on Windows:

Issue is with input_shape. input_shape=input_shape[1:]

Working sample code

You need to add [0] as indexing,

so where you wrote self.logger.record('reward', self.training_env.get_attr('total_reward')) you just need to index with self.logger.record('reward', self.training_env.get_attr ('total_reward')[0])

Essentially, what happens here is that the output of the net is being sliced to get the desired part of the Q table.

The (somewhat confusing) index of [np.arange(0, self.batch_size), action] indexes each axis. So, for axis with index 1, we pick the item indicated by action. For index 0, we pick all items between 0 and self.batch_size.

If self.batch_size is the same as the length of dimension 0 of this array, then this slice can be simplified to [:, action] which is probably more familiar to most users.

There was nothing wrong with the network definition. It turns out the learning rate was too high and reducing it 0.00025 (as in the original Nature paper introducing the DQN) led to an agent which can solve CartPole-v0.

That said, the learning algorithm was incorrect. In particular I was using the wrong target action-value predictions. Note the algorithm laid out above does not use the most recent version of the target network to make predictions. This leads to poor results as training progresses because the agent is learning based on stale target data. The way to fix this is to just put (s, a, r, s', done) into the replay memory and then make target predictions using the most up to date version of the target network when sampling a mini batch. See the code below for an updated learning loop.