I'm running a SAC reinforcement learner for a robotics application with some pretty decent results. One of the reasons I opted for reinforcement learning is for the ability for learning in the field, e.g. to adjust to a mechanical change, such as worn tires or a wheel going a little out of alignment.
My reinforcement learner restores it's last saved weights and replay buffer upon startup, so it doesn't need to retrain every time I turn it on. However, one concern I have is with respect to the optimizer.
Optimizers have come a long way since ADAM, but everything I read and all the RL code samples I see still seem to use ADAM with a fixed learning rate. I'd like to take advantage of some of the advances in optimizers, e.g. one cycle AdamW. However, a one-cycle optimizer seems inappropriate for a continuous real-world reinforcement learning problem: I imagine it's pretty good for the initial training/calibration, but I expect the low final learning rate would react too slowly to mechanical changes.
One thought I had was perhaps to do a one-cycle approach for initial training, and triggering a smaller one-cycle restart if a change in error that indicates something has changed (perhaps the size of the restart could be based on the size of the change in error).
Has anyone experimented with optimizers other than ADAM for reinforcement learning or have any suggestions for dealing with this sort of problem?
Reinforcement learning is very different from traditional supervised learning because the training data distribution changes as the policy improves. In optimization terms, the objective function can be said to be non-stationary. For this reason, I suspect your intuition is likely correct -- that a "one-cycle" optimizer would perform poorly after a while in your application.
My question is, what is wrong with Adam? Typically, the choice of optimizer is a minor detail for deep reinforcement learning; other factors like the exploration policy, algorithmic hyperparameters, or network architecture tend to have a much greater impact on performance.
Nevertheless, if you really want to try other optimizers, you could experiment with RMSProp, Adadelta, or Nesterov Momentum. However, my guess is that you will see incremental improvements, if any. Perhaps searching for better hyperparameters to use with Adam would be a more effective use of time.
EDIT: In my original answer, I made the claim that the choice of a particular optimizer is not primarily important for reinforcement learning speed, and neither is generalization. I want to add some discussion that helps illustrate these points.
Consider how most deep policy gradient methods operate: they sample a trajectory of experience from the environment, estimate returns, and then conduct one or more gradient steps to improve the parameterized policy (e.g. a neural network). This process repeats until convergence (to a locally optimal policy).
Why must we continuously sample new experience from the environment? Because our current data can only provide a reasonable first-order approximation within a small trust region around the policy parameters that were used to collect that data. Hence, whenever we update the policy, we need to sample more data.
A good way to visualize this is to consider an MM algorithm. At each iteration, a surrogate objective is constructed based on the data we have now and then maximized. Each time, we will get closer to the true optimum, but the speed at which we approach it is determined only by the number of surrogates we construct -- not by the specific optimizer we use to maximize each surrogate. Adam might maximize each surrogate in fewer gradient steps than, say, RMSProp does, but this does not affect the learning speed of the agent (with respect to environment samples). It just reduces the number of minibatch updates you need to conduct.
SAC is a little more complicated than this, as it learns Q-values in an off-policy manner and conducts updates using experience replay, but the general idea holds. The best attainable policy is subject to whatever the current data in our replay memory are; regardless of the optimizer we use, we will need to sample roughly the same amount of data from the environment to converge to the optimal policy.
So, how do you make a faster (more sample-efficient) policy gradient method? You need to fundamentally change the RL algorithm itself. For example, PPO almost always learns faster than TRPO, because John Schulman and co-authors found a different and empirically better way to generate policy gradient steps.
Finally, notice that there is no notion of generalization here. We have an objective function that we want to optimize, and once we do optimize it, we have solved the task as well as we can. This is why I suspect that the "Adam-generalizes-worse-than-SGD" issue is actually irrelevant for RL.
My initial testing suggest the details of the optimizer and it's hyperparameters matter, at least for off-policy techniques. I haven't had the chance to experiment much with PPO or on-policy techniques, so I can't speak for those unfortunately.
To speak to #Brett_Daley's thoughtful response a bit: the optimizer is certainly one of the less important characteristics. The means of exploration, and the use of a good prioritized replay buffer are certainly critical factors, especially with respect to achieving good initial results. However, my testing seems to show that the optimizer becomes important for the fine-tuning.
The off-policy methods I have been using have been problematic with fine-grained stability. In other words, the RL finds the mostly correct solution, but never really hones in on the perfect solution (or if it does find it briefly, it drifts off). I suspect the optimizer is at least partly to blame.
I did a bit of testing and found that varying the ADAM learning rate has an obvious effect. Too high and both the actor and critic bounce around the minimum and never converge on the optimal policy. In my robotics application this looks like the RL consistently makes sub-optimal decisions, as though there's a bit of random exploration with every action that always misses the mark a little bit.
OTOH, a lower learning rate tends to get stuck in sub-optimal solutions and is unable to adapt to changes (e.g. slower motor response due to low battery).
I haven't yet run any tests of single-cycle schedule or AdamW for the learning rate, but I did a very basic test with a two stage learning rate adjustment for both Actor and Critic (starting with a high rate and dropping to a low rate) and the results were a clearly more precise solution that converged quickly during the high learning rate and then honed in better with the low-learning rate.
I imagine AdamW's better weight decay regularization may result in similarly better results for avoiding overfitting training batches contributing to missing the optimal solution.
Based on the improvement I saw, it's probably worth trying single-cycle methods and AdamW for the actor and critic networks for tuning the results. I still have some concerns for how the lower learning rate at the end of the cycle will adapt to changes in the environment, but a simple solution for that may be to monitor the loss and do a restart of the learning rate if it drifts too much. In any case, more testing seems warranted.
I want to use DQN on recommendation system for retail industry
but the problem is, the state space of this question are time-inhomogeneous & not deterministic
(compare to Atari games)
I figure out two method for this problem
make state-transition become deterministic
use historical data to calculate transition probabilities, use probabilities to transit state
but...both of them seems not make sense
somebody point out this kind issues
if I want to build a recommendation system based on Reinforcement Learning
where should I start?
I have little background knowledge of Machine Learning, so please forgive me if my question seems silly.
Based on what I've read, the best model-free reinforcement learning algorithm to this date is Q-Learning, where each state,action pair in the agent's world is given a q-value, and at each state the action with the highest q-value is chosen. The q-value is then updated as follows:
Q(s,a) = (1-α)Q(s,a) + α(R(s,a,s') + (max_a' * Q(s',a'))) where α is the learning rate.
Apparently, for problems with high dimensionality, the number of states become astronomically large making q-value table storage infeasible.
So the practical implementation of Q-Learning requires using Q-value approximation via generalization of states aka features. For example if the agent was Pacman then the features would be:
Distance to closest dot
Distance to closest ghost
Is Pacman in a tunnel?
And then instead of q-values for every single state you would only need to only have q-values for every single feature.
So my question is:
Is it possible for a reinforcement learning agent to create or generate additional features?
Some research I've done:
This post mentions A Geramifard's iFDD method
http://www.icml-2011.org/papers/473_icmlpaper.pdf
http://people.csail.mit.edu/agf/Files/13RLDM-GQ-iFDD+.pdf
which is a way of "discovering feature dependencies", but I'm not sure if that is feature generation, as the paper assumes that you start off with a set of binary features.
Another paper that I found was apropos is Playing Atari with Deep Reinforcement Learning, which "extracts high level features using a range of neural network architectures".
I've read over the paper but still need to flesh out/fully understand their algorithm. Is this what I'm looking for?
Thanks
It seems like you already answered your own question :)
Feature generation is not part of the Q-learning (and SARSA) algorithm. In a process which is called preprocessing you can however use a wide array of algorithms (of which you showed some) to generate/extract features from your data. Combining different machine learning algorithms results in hybrid architectures, which is a term you might look into when researching what works best for your problem.
Here is an example of using features with SARSA (which is very similar to Q-learning).
Whether the papers you cited are helpful for your scenario, you'll have to decide for yourself. As always with machine learning, your approach is highly problem-dependent. If you're in robotics and it's hard to define discrete states manually, a neural network might be helpful. If you can think of heuristics by yourself (like in the pacman example) then you probably won't need it.
I have a minimax game playing program that sums together different heuristics to return a value for each state of the game. I would like to implement learning. I want the program to learn weights for each heuristic. What is the most effective means of having the program learn the weights for each heuristic? Of course, it would only know if a certain weight was effective for a certain heuristic after trying it. Is the only option some kind of trial and error system?
Thank you for your help!
I've not applied minimax much in practice - but in general its preferable to have an intrinsic measure of score/goodness/badness to base it off of. The first step would be to try and define such a score for you game - and expose that as an interface that is implemented for each supported game.
Is the only option some kind of trial and error system?
No! Genetic algorithms are popular for this kind of thing (at least among hobbyists), and can be used successfully for many problems (given sufficient time). You can find a lot of information related to this in early AI research, especially related to chess programs.
You can look up some of the research in hyperparameter optimization to find more machine learning style ways to do it. Unfortunately its not as well studied an area as it probably should be.
There are more possibilities depending on the specifics of the game being implemented / the nature of the heuristics.
Reinforcement Learning (RL), in particual Temporal Difference (TD) methods, deal with learning weights for heuristics in non-adversarial setting. How to learn weights for heuristics in a game setting, depends on what algorithms you use to play the game. The major classes of algorithms are alpha-beta minimax and UpperConfidenceTree. For minimax, you can look at the updates to values on the tree nodes as you increase the depth of the tree. I recommend starting by learning about RL-TD, and then reading Bootstrapping from Game Tree Search
by Joel Veness et. al.
Closed. This question is off-topic. It is not currently accepting answers.
Want to improve this question? Update the question so it's on-topic for Stack Overflow.
Closed 10 years ago.
Improve this question
What is machine learning ?
What does machine learning code do ?
When we say that the machine learns, does it modify the code of itself or it modifies history (database) which will contain the experience of code for given set of inputs?
What is a machine learning ?
Essentially, it is a method of teaching computers to make and improve predictions or behaviors based on some data. What is this "data"? Well, that depends entirely on the problem. It could be readings from a robot's sensors as it learns to walk, or the correct output of a program for certain input.
Another way to think about machine learning is that it is "pattern recognition" - the act of teaching a program to react to or recognize patterns.
What does machine learning code do ?
Depends on the type of machine learning you're talking about. Machine learning is a huge field, with hundreds of different algorithms for solving myriad different problems - see Wikipedia for more information; specifically, look under Algorithm Types.
When we say machine learns, does it modify the code of itself or it modifies history (Data Base) which will contain the experience of code for given set of inputs ?
Once again, it depends.
One example of code actually being modified is Genetic Programming, where you essentially evolve a program to complete a task (of course, the program doesn't modify itself - but it does modify another computer program).
Neural networks, on the other hand, modify their parameters automatically in response to prepared stimuli and expected response. This allows them to produce many behaviors (theoretically, they can produce any behavior because they can approximate any function to an arbitrary precision, given enough time).
I should note that your use of the term "database" implies that machine learning algorithms work by "remembering" information, events, or experiences. This is not necessarily (or even often!) the case.
Neural networks, which I already mentioned, only keep the current "state" of the approximation, which is updated as learning occurs. Rather than remembering what happened and how to react to it, neural networks build a sort of "model" of their "world." The model tells them how to react to certain inputs, even if the inputs are something that it has never seen before.
This last ability - the ability to react to inputs that have never been seen before - is one of the core tenets of many machine learning algorithms. Imagine trying to teach a computer driver to navigate highways in traffic. Using your "database" metaphor, you would have to teach the computer exactly what to do in millions of possible situations. An effective machine learning algorithm would (hopefully!) be able to learn similarities between different states and react to them similarly.
The similarities between states can be anything - even things we might think of as "mundane" can really trip up a computer! For example, let's say that the computer driver learned that when a car in front of it slowed down, it had to slow down to. For a human, replacing the car with a motorcycle doesn't change anything - we recognize that the motorcycle is also a vehicle. For a machine learning algorithm, this can actually be surprisingly difficult! A database would have to store information separately about the case where a car is in front and where a motorcycle is in front. A machine learning algorithm, on the other hand, would "learn" from the car example and be able to generalize to the motorcycle example automatically.
Machine learning is a field of computer science, probability theory, and optimization theory which allows complex tasks to be solved for which a logical/procedural approach would not be possible or feasible.
There are several different categories of machine learning, including (but not limited to):
Supervised learning
Reinforcement learning
Supervised Learning
In supervised learning, you have some really complex function (mapping) from inputs to outputs, you have lots of examples of input/output pairs, but you don't know what that complicated function is. A supervised learning algorithm makes it possible, given a large data set of input/output pairs, to predict the output value for some new input value that you may not have seen before. The basic method is that you break the data set down into a training set and a test set. You have some model with an associated error function which you try to minimize over the training set, and then you make sure that your solution works on the test set. Once you have repeated this with different machine learning algorithms and/or parameters until the model performs reasonably well on the test set, then you can attempt to use the result on new inputs. Note that in this case, the program does not change, only the model (data) is changed. Although one could, theoretically, output a different program, but that is not done in practice, as far as I am aware. An example of supervised learning would be the digit recognition system used by the post office, where it maps the pixels to labels in the set 0...9, using a large set of pictures of digits that were labeled by hand as being in 0...9.
Reinforcement Learning
In reinforcement learning, the program is responsible for making decisions, and it periodically receives some sort of award/utility for its actions. However, unlike in the supervised learning case, the results are not immediate; the algorithm could prescribe a large sequence of actions and only receive feedback at the very end. In reinforcement learning, the goal is to build up a good model such that the algorithm will generate the sequence of decisions that lead to the highest long term utility/reward. A good example of reinforcement learning is teaching a robot how to navigate by giving a negative penalty whenever its bump sensor detects that it has bumped into an object. If coded correctly, it is possible for the robot to eventually correlate its range finder sensor data with its bumper sensor data and the directions that sends to the wheels, and ultimately choose a form of navigation that results in it not bumping into objects.
More Info
If you are interested in learning more, I strongly recommend that you read Pattern Recognition and Machine Learning by Christopher M. Bishop or take a machine learning course. You may also be interested in reading, for free, the lecture notes from CIS 520: Machine Learning at Penn.
Machine learning is a scientific discipline that is concerned with the design and development of algorithms that allow computers to evolve behaviors based on empirical data, such as from sensor data or databases. Read more on Wikipedia
Machine learning code records "facts" or approximations in some sort of storage, and with the algorithms calculates different probabilities.
The code itself will not be modified when a machine learns, only the database of what "it knows".
Machine learning is a methodology to create a model based on sample data and use the model to make a prediction or strategy. It belongs to artificial intelligence.
Machine learning is simply a generic term to define a variety of learning algorithms that produce a quasi learning from examples (unlabeled/labeled). The actual accuracy/error is entirely determined by the quality of training/test data you provide to your learning algorithm. This can be measured using a convergence rate. The reason you provide examples is because you want the learning algorithm of your choice to be able to informatively by guidance make generalization. The algorithms can be classed into two main areas supervised learning(classification) and unsupervised learning(clustering) techniques. It is extremely important that you make an informed decision on how you plan on separating your training and test data sets as well as the quality that you provide to your learning algorithm. When you providing data sets you want to also be aware of things like over fitting and maintaining a sense of healthy bias in your examples. The algorithm then basically learns wrote to wrote on the basis of generalization it achieves from the data you have provided to it both for training and then for testing in process you try to get your learning algorithm to produce new examples on basis of your targeted training. In clustering there is very little informative guidance the algorithm basically tries to produce through measures of patterns between data to build related sets of clusters e.g kmeans/knearest neighbor.
some good books:
Introduction to ML (Nilsson/Stanford),
Gaussian Process for ML,
Introduction to ML (Alpaydin),
Information Theory Inference and Learning Algorithms (very useful book),
Machine Learning (Mitchell),
Pattern Recognition and Machine Learning (standard ML course book at Edinburgh and various Unis but relatively a heavy reading with math),
Data Mining and Practical Machine Learning with Weka (work through the theory using weka and practice in Java)
Reinforcement Learning there is a free book online you can read:
http://www.cs.ualberta.ca/~sutton/book/ebook/the-book.html
IR, IE, Recommenders, and Text/Data/Web Mining in general use alot of Machine Learning principles. You can even apply Metaheuristic/Global Optimization Techniques here to further automate your learning processes. e.g apply an evolutionary technique like GA (genetic algorithm) to optimize your neural network based approach (which may use some learning algorithm). You can approach it purely in form of a probablistic machine learning approach for example bayesian learning. Most of these algorithms all have a very heavy use of statistics. Concepts of convergence and generalization are important to many of these learning algorithms.
Machine learning is the study in computing science of making algorithms that are able to classify information they haven't seen before, by learning patterns from training on similar information. There are all sorts of kinds of "learners" in this sense. Neural networks, Bayesian networks, decision trees, k-clustering algorithms, hidden markov models and support vector machines are examples.
Based on the learner, they each learn in different ways. Some learners produce human-understandable frameworks (e.g. decision trees), and some are generally inscrutable (e.g. neural networks).
Learners are all essentially data-driven, meaning they save their state as data to be reused later. They aren't self-modifying as such, at least in general.
I think one of the coolest definitions of machine learning that I've read is from this book by Tom Mitchell. Easy to remember and intuitive.
A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E
Shamelessly ripped from Wikipedia: Machine learning is a scientific discipline that is concerned with the design and development of algorithms that allow computers to evolve behaviors based on empirical data, such as from sensor data or databases.
Quite simply, machine learning code accomplishes a machine learning task. That can be a number of things from interpreting sensor data to a genetic algorithm.
I would say it depends. No, modifying code is not normal, but is not outside the realm of possibility. I would also not say that machine learning always modifies a history. Sometimes we have no history to build off of. Sometime we simply want to react to the environment, but not actually learn from our past experiences.
Basically, machine learning is a very wide-open discipline that contains many methods and algorithms that make it impossible for there to be 1 answer to your 3rd question.
Machine learning is a term that is taken from the real world of a person, and applied on something that can't actually learn - a machine.
To add to the other answers - machine learning will not (usually) change the code, but it might change it's execution path and decision based on previous data or new gathered data and hence the "learning" effect.
there are many ways to "teach" a machine - you give weights to many parameter of an algorithm, and then have the machine solve it for many cases, each time you give her a feedback about the answer and the machine adjusts the weights according to how close the machine answer was to your answer or according to the score you gave it's answer, or according to some results test algorithm.
This is one way of learning and there are many more...