15
How does this method differ from the experience replay, as they both use past information in the training? What's the typical application of both techniques?
Using a recurrent neural network is one way for an agent to build a model of hidden or unobserved state in order to improve its predictions when direct observations do not give enough information, but ...
13
What you are looking for is called "reinforcement learning".
A reinforcement learning algorithm will try to maximize a reward function. This reward represents how "good" or "bad" an action is in the actual context. For example, in the snake game, your reward will be positive for eating an apple and negative when the snake hits a ...
12
Here's an intuitive description answer:
Function approximation can be done with any parameterizable function. Consider the problem of a $Q(s,a)$ space where $s$ is the positive reals, $a$ is $0$ or $1$, and the true Q-function is $Q(s, 0) = s^2$, and $Q(s, 1)= 2s^2$, for all states. If your function approximator is $Q(s, a) = m*s + n*a + b$, there exists no ...
10
My best guess that it's been done to reduce the computation time, otherwise we would have to find out the q value for each action and then select the best one.
It has no real impact on computation time, other than a slight increase (due to extra memory used by two networks). You could cache results of the target network I suppose, but it probably would not ...
9
Here is a table that attempts to systematically show the differences between tabular Q-learning (TQL), deep Q-learning (DQL), and deep Q-network (DQN).
Tabular Q-learning (TQL)
Deep Q-learning (DQL)
Deep Q-network (DQN)
Is it an RL algorithm?
Yes
Yes
No (unless you use DQN to refer to DQL, which is done often!)
Does it use neural networks?
No. It uses a ...
8
Gradient descent and back-propagation
In deep learning, gradient descent (GD) and back-propagation (BP) are used to update the weights of the neural network.
In reinforcement learning, one could map (state, action)-pairs to Q-values with a neural network. Then GD and BP can be used to update the weights of this neural network.
How to design the neural ...
8
The speaker argued that a dream is a random addition of memories, just as experience replay.
The speaker is taking some liberties due to a general lack of scientific understanding of what dreams are. We don't even have strong consensus on why sleep is a necessary feature of animals, let alone what part dreaming plays in it. However, there are some widely-...
7
Eligibility traces is a method of weighting between temporal-difference "targets" and Monte-Carlo "returns". In practice, for example, instead of using the one-step TD target, $r_t + \gamma V (s_{t+1})$, as in the temporal difference update $V (s_t) \leftarrow V (s_t) + \alpha (r_t + \gamma V (s_{t+1}) − V (s_t))$, you use the so-called &...
7
No. DQN and other deep RL methods work well with fully connected layers. Here's an implementation of DQN which doesn't use CNNs: github.com/keon/deep-q-learning/blob/master/dqn.py
DeepMind mostly use CNN because they use image as input state, and that because they tried to evaluate performance of their methods vs humans performance. Humane performance is ...
7
There is a relatively recent paper that tackles this issue: Challenges of real-world reinforcement learning (2019) by Gabriel Dulac-Arnold et al., which presents all the challenges that need to be addressed to productionize RL to real world problems, the current approaches/solutions to solve the challenges, and metrics to evaluate them. I will only list them ...
6
To model chess as a Markov decision problem (MDP) you can refer to the AlphaZero paper (Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm). The exact details can be found starting from the bottom of page 13.
Briefly, an action is described by picking a piece and then picking a move with it. The size of the board is 8 by ...
6
DQN and AlphaZero do not share much in terms of implementation.
However, they are based on the same Reinforcement Learning (RL) theoretical framework. If you understand terms like MDP, reward, return, value, policy, then these are interchangeable between DQN and AlphaZero. When it comes to implementation, and what each part of the system is doing, then this ...
6
I have done some research and would like to share.
Generally to eliminate the use of target network one needs to show that training would be stable under off-policy semi-gradient.
There are two approaches that might work:
Experience reweighting
Constrained optimization
Experience reweighting
Probably the simplest idea is to use importance sampling ...
6
In order for the algorithm to have stable behavior, the replay buffer should be large enough to contain a wide range of experiences, but it may not always be good to keep everything.
The larger the experience replay, the less likely you will sample correlated elements, hence the more stable the training of the NN will be. However, a large experience replay ...
6
$Q$-learning (and also its deep variant, and most of the other well-known reinforcement learning algorithms) are inherently learning approaches for single-agent environments. The entire problem setting that these algorithms are developed for (Markov decision processes, or MDPs) is always framed in terms of a single agent situated in some environment, where ...
5
We can start with equation (30):
$$
\bar{A}(s) = P(a \neq \tilde{a}) \mathbb{E}_{(a,\tilde{a})\sim(\pi,\tilde{\pi}|a\neq\tilde{a})} [A_\pi(s, \tilde{a}) - A_\pi(s, a)]
$$
Taking the absolute value of both sides, the equality remains true. We can pull the probability term out of the absolute value since it is guaranteed to be nonnegative.
$$
|\bar{A}(s)| = ...
5
I am not 100% sure if the following is the only/complete story, but I'm quite confident it's at least part of the story:
In the PPO paper, after describing the standard policy gradient objective $L^{PG}$, they mention the following:
While it is appealing to perform multiple steps of optimization on this loss $L^{PG}$ using the same trajectory, doing so ...
5
Instead of having the AI learn what action to take, you can alternatively train it to judge how "good" a position is. In order to determine what move to make, you don't ask the AI "This is the current state, what move should I make", you iterate through all possible moves, and feed the the resulting state into the AI asking "How good do you think this new ...
5
You need to read this 2020 paper by Deepmind:
"Revisiting Fundamentals of Experience Replay"
Also, to add to the answer by @nbro
Assume you implement experience replay as a buffer where the newest memory is stored instead of the oldest. Then, if your buffer contains 100k entries, any memory will remain there for exactly 100k iterations.
Such a ...
5
As far as I'm aware, it is still somewhat of an open problem to get a really clear, formal understanding of exactly why / when we get a lack of convergence -- or, worse, sometimes a danger of divergence. It is typically attributed to the "deadly triad" (see 11.3 of the second edition of Sutton and Barto's book), the combination of:
Function approximation, ...
5
This answer will point the reader to potentially useful resources, but I can't ensure that the courses are good (because I have never followed them).
Free
Reinforcement Learning in the Open AI Gym (a small course that you can find in the YouTube channel suggested in the other answer) by Phil Tabor
The free course Advanced Deep Learning & Reinforcement ...
5
If you're interested in the theory behind Double Q-learning (not deep!), the reference paper would be Double Q-learning by Hado van Hasselt (2010).
As for Double deep Q-learning (also called DDQN, short for Double Deep Q-networks), the reference paper would be Deep Reinforcement Learning with Double Q-learning by Van Hasselt et al. (2016), as pointed out ...
5
In Q-learning (and in general value based reinforcement learning) we are typically interested in learning a Q-function, $Q(s, a)$. This is defined as
$$Q(s, a) = \mathbb{E}_\pi\left[ G_t | S_t = s, A_t = a \right]\;.$$
For tabular Q-learning, where you have a finite state and action space - note that in practice even the spaces being finite might not be ...
4
Filling values is totally fine. In the case of image recognition the filling will be the background of the image (examples). For example in Belot you have total of 32 cards, which can be 32 boolean features. You can set the ones the player has to 1, while the rest are 0. Note that the in most games you'll need more features than the cards in your hand. I.e ...
4
After some research and reading this post, I see where my problem was: I was introducing a full consecutive batch of experiences, selected randomly, yes, but the experiences in the batch were consecutives.
After redoing my experience selection method, my DQN is actually working and has reached about +200 points after 400000 experiences (about 500 episodes; ...
4
I think this was just a "clever" design choice. You can actually design a neural network (NN), to represent your Q function, which receives as input the state and an action and outputs the corresponding Q value. However, to obtain $\max_aQ(s', a)$ (which is a term of the update rule of the Q-learning algorithm) you would need a "forward pass" of this network ...
4
Even after several years of success of deep learning systems (i.e. neural networks trained with gradient descent and back-propagation), as far as I know, there is not yet a consensus on what constitutes a deep neural network. Some people could use a neural network with 2 hidden layers and call it deep (like in your case), but other people may just dedicate ...
4
In some cases we may wish to have a discount factor $\gamma_t$ which depends on time $t$ (or depends on state $s_t$ and/or action $a_t$, leading to an indirect dependence on time $t$). Indeed we do not usually do this, but it does happen sometimes.
I guess that, from a theoretical point of view, it was very easy of the authors to make their algorithm more ...
4
First, the derivative is usually taken with respect to a variable (input) of the function. Hence the notation $\frac{df}{dx}$ for some function $f(x)$.
If you look at your equation more carefully
$$\nabla log P(\tau^{i};\theta) = \Sigma_{t=0}\nabla_{\theta}log\pi(a_{t}|s_t, \theta).$$
You will see that the gradient is taken with respect to $\theta$, which ...
4
For the programming part I suggest this YouTube channel by Phil Tabor (he also has a website: neuralnet.ai. I found his videos really useful while I was attending reinforcement learning classes at the uni. He covers basic algorithms like value iteration and policy iteration and also more advanced like deep q learning, covering all main python libraries (...
Only top voted, non community-wiki answers of a minimum length are eligible
