The authors propose a unified setting to evaluate the performance of batch reinforcement learning algorithms. The proposed benchmark is discrete and based on the popular Atari Domain. The authors review and benchmark several current batch RL algorithms against a newly introduced version of BCQ (Batch Constrained Deep Q Learning) for discrete environments. 

https://i.imgur.com/zrCZ173.png

Note in line 5 that the policy chooses actions with a restricted argmax operation, eliminating actions that have not enough support in the batch. 

One of the key difficulties in batch-RL is the divergence of value estimates. In this paper the authors use Double DQN, which means actions are selected with a value net $Q_{\theta}$ and the policy evaluation is done with a target network $Q_{\theta'}$ (line 6). 

**How is the batch created?**
A partially trained DQN-agent (trained online for 10mio steps, aka 40mio frames) is used as behavioral policy to collect a batch $B$ containing 10mio transitions. The DQN agent uses either with probability 0.8 an $\epsilon=0.2$ and with probability 0.2 an $\epsilon = 0.001$. The batch RL agents are trained on this batch for 10mio steps and evaluated every 50k time steps for 10 episodes. This process of batch creation differs from the settings used in other papers in i) having only a single behavioral policy, ii) the batch size and iii) the proficiency level of the batch policy.

The experiments, performed on the arcade learning environment include DQN, REM, QR-DQN, KL-Control, BCQ, OnlineDQN and Behavioral Cloning and show that:
- for conventional RL algorithms distributional algorithms (QR-DQN) outperform the plain algorithms (DQN)
- batch RL algorithms perform better than conventional algorithms with BCQ outperforming every other algorithm in every tested game

In addition to the return the authors plot the value estimates for the Q-networks. A drop in performance corresponds in all cases to a divergence (up or down) in value estimates. 

The paper is an important contribution to the debate about what is the right setting to evaluate batch RL algorithms. It remains however to be seen if the proposed choice of i) a single behavior policy, ii) the batch size and iii) quality level of the behavior policy will be accepted as standard. Further work is in any case required to decide upon a benchmark for continuous domains.

Interacting with the environment comes sometimes at a high cost, for example in high stake scenarios like health care or teaching. Thus instead of learning online, we might want to learn from a fixed buffer $B$ of transitions, which is filled in advance from a behavior policy.
The authors show that several so called off-policy algorithms, like DQN and DDPG fail dramatically in this pure off-policy setting. 

They attribute this to the extrapolation error, which occurs in the update of a value estimate $Q(s,a)$, where the target policy selects an unfamiliar action $\pi(s')$ such that $(s', \pi(s'))$ is unlikely or not present in $B$. Extrapolation error is caused by the mismatch between the true state-action visitation distribution of the current policy and the state-action distribution in $B$ due to:
- state-action pairs (s,a) missing  in $B$, resulting in arbitrarily bad estimates of $Q_{\theta}(s, a)$ without sufficient data close to (s,a). 
- the finiteness of the batch of transition tuples $B$, leading to a biased estimate of the transition dynamics in the Bellman operator $T^{\pi}Q(s,a) \approx \mathbb{E}_{\boldsymbol{s' \sim B}}\left[r + \gamma Q(s', \pi(s')) \right]$
- transitions are sampled uniformly from $B$, resulting in a loss weighted w.r.t the frequency of data in the batch: $\frac{1}{\vert B \vert} \sum_{\boldsymbol{(s, a, r, s') \sim B}} \Vert r + \gamma Q(s', \pi(s')) - Q(s, a)\Vert^2$

The proposed algorithm Batch-Constrained deep Q-learning (BCQ) aims to choose actions that:
 1. minimize distance of taken actions to actions in the batch
 2. lead to states contained in the buffer
 3.  maximizes the value function,

where 1. is prioritized over the other two goals to mitigate the extrapolation error. 

Their proposed algorithm (for continuous environments) consists informally of the following steps that are repeated at each time $t$:
 1. update generator model of the state conditional marginal likelihood $P_B^G(a \vert s)$
 2. sample n actions form the generator model
 3. perturb each of the sampled actions to lie in a range $\left[-\Phi, \Phi \right]$
 4. act according to the argmax of respective Q-values of perturbed actions
 5. update value function

The experiments considers Mujoco tasks with four scenarios of batch data creation:
 - 1 million time steps from training a DDPG agent with exploration noise $\mathcal{N}(0,0.5)$ added to the action.This aims for a diverse set of states and actions. 
 - 1 million time steps from training a DDPG agent with an exploration noise $\mathcal{N}(0,0.1)$ added to the actions as behavior policy. The batch-RL agent and the behavior DDPG are trained concurrently from the same buffer. 
 - 1 million transitions from rolling out a already trained DDPG agent
 - 100k transitions from a behavior policy that acts with probability 0.3 randomly and follows otherwise an expert demonstration with added exploration noise  $\mathcal{N}(0,0.3)$

I like the fourth choice of behavior policy the most as this captures high stake scenarios like education or medicine the closest, in which training data would be acquired by human experts that are by the nature of humans not optimal but significantly better than learning from scratch. 

The proposed BCQ algorithm is the only algorithm that is successful across all experiments. It matches or outperforms the behavior policy. Evaluation of the value estimates showcases unstable and diverging value estimates for all algorithms but BCQ that exhibits a stable value function. 

The paper outlines a very important issue that needs to be tackled in order to use reinforcement learning in real world applications. 

disclaimer: I'm the first author of the paper

## TL;DR 

We have made a lot of progress on catastrophic forgetting within the standard evaluation protocol,
i.e. sequentially learning a stream of tasks and testing our models' capacity to remember them all.

We think it's time a new approach to Continual Learning (CL), coined OSAKA, which is more aligned with real-life applications of CL. It brings CL closer to Online Learning and Open-World Learning.

main modifications we propose:
- bring CL closer to Online learning i.e. at test time, the model is continually learning and evaluated on its online predictions 
- it's fine to forget, as long as you can quickly remember (just like we humans do)
- we allow pretraining, (because you wouldn't deploy an untrained CL system, right?) but at test time, the model will have to quickly learn new out-of-distribution (OoD) tasks (because the world is full of surprises)
- the tasks distribution is actually a hidden Markov chain. This implies:
    - new and old tasks can re-occur (just like in real life). Better remember them quickly if you want to get a good total performance!
    - tasks have different lengths
    - and the tasks boundaries are unknown (task agnostic setting)

### Bonus:
We provide a unifying framework explaining the space of machine learning setting {supervised learning, meta learning, continual learning, meta-continual learning, continual-meta learning} in case it was starting to get confusing :p

## Motivation

We imagine an agent, embedded or not, first pre-trained in a controlled environment and later deployed in the real world, where it faces new or unexpected situations. This scenario is relevant for many applications. For instance, in robotics, the agent is pre-trained in a factory and deployed in homes or
in manufactures where it will need to adapt to new domains and maybe solve new tasks. Likewise, a virtual assistant can be pre-trained on static datasets and deployed in a user’s life to fit its personal needs. 

Further motivations can be found in time series forecasting, e.g., market prediction,
game playing, autonomous customer service, recommendation systems, autonomous driving, to name a few. In this scenario, we are interested in the cumulative performance of the agent throughout its lifetime. Differently, standard CL reports the agent’s final performance on all tasks at the end of its life. In order
to succeed in this scenario, agents need the ability to learn new tasks as well as quickly remembering old ones.

## Unifying Framework

We propose a unifying framework explaining the space of machine learning setting {supervised learning, meta learning, continual learning, meta-continual learning, continual-meta learning} with meta learning terminology. 

https://i.imgur.com/U16kHXk.png

(easier to digest with accompanying text)

## OSAKA

The main features of the evaluation framework are
- task agnosticism
- pre-training is allowed, but OoD tasks at test time
- task revisiting
- controllable non-stationarity
- online evaluation

(see paper for the motivations of the features)

## Continual-MAML: an initial baseline

A simple extension of MAML that is better suited than previous methods in the proposed setting.

https://i.imgur.com/C86WUc8.png

Features are:
- Fast Adapatation
- Dynamic Representation
- Task boundary detection
- Computational efficiency


## Experiments

We provide a suite of 3 benchmarks to test algorithms in the new setting.

The first includes the Omniglot, MNIST and FashionMNIST dataset. 

The second and third use the Synbols (Lacoste et al. 2018) and TieredImageNet datasets, respectively.

The first set of experiments shows that the baseline outperforms previous approaches, i.e., supervised learning, meta learning, continual learning, meta-continual learning, continual-meta learning, in the new setting.

https://i.imgur.com/IQ1WYTp.png

The second and third experiments lead us to similar conclusions

code: https://github.com/ElementAI/osaka

The authors analyse in the very well written paper the relation between Fisher $F(\theta) = \sum_n \mathbb{E}_{p_{\theta}(y \vert x)}[\nabla_{\theta} \log(p_{\theta}(y \vert x_n))\nabla_{\theta} \log(p_{\theta}(y \vert x_n))^T] $ and empirical Fisher $\bar{F}(\theta) = \sum_n [\nabla_{\theta} \log(p_{\theta}(y_n \vert x_n))\nabla_{\theta} \log(p_{\theta}(y_n \vert x_n))^T] $, which has recently seen a surge in interest. . The definitions differ in that $y_n$ is a training label instead of a sample of the model $p_{\theta}(y \vert x_n)$, thus even so the name suggests otherwise $\bar{F}$ is not a empirical, for example Monte Carlo, estimate of the Fisher. The authors rebuff common arguments used to justify the use of the empirical fisher by an amendment to the generalized Gauss-Newton, give conditions when the empirical Fisher does indeed approach the Fisher and give an argument why the empirical fisher might work in practice nonetheless.   

The Fisher, capturing the curvature of the parameter space, provides information about the geometry of the parameters pace, the empirical Fisher might however fail so capture the curvature as the striking plot from the paper shows:
https://i.imgur.com/c5iCqXW.png

The authors rebuff the two major justifications for the use of empirical Fisher:
1. "the empirical Fisher matches the construction of a generalized Gauss-Newton"
 * for the log-likelihood $log(p(y \vert f) = \log \exp(-\frac{1}{2}(y-f)^2))$ the generalized Gauss-Newton intuition that small residuals $f(x_n, \theta) - y_n$ lead to a good approximation of the Hessian is not satisfied. Whereas the Fisher approaches the Hessian, the empirical Fisher approaches 0
2. "the empirical Fisher converges to the true Fisher when the model is a good fit for the data"
 * the authors sharpen the argument to "the empirical Fisher converges at the minimum to the Fisher as the number of samples grows", which is unlikely to be satisfied in practice. 

The authors provide an alternative perspective on why the empirical Fisher might be successful, namely to adapt the gradient to the gradient noise in stochastic optimization. The empirical Fisher coincides with the second moment of the stochastic gradient estimate and encodes as such covariance information about the gradient noise. This allows to reduce the effects of gradient noise by scaling back the updates in high variance aka noise directions. 

Bafna et al. show that iterative hard thresholding results in $L_0$ robust Fourier transforms. In particular, as shown in Algorithm 1, iterative hard thresholding assumes a signal $y = x + e$ where $x$ is assumed to be sparse, and $e$ is assumed to be sparse. This translates to noise $e$ that is bounded in its $L_0$ norm, corresponding to common adversarial attacks such as adversarial patches in computer vision. Using their algorithm, the authors can provably reconstruct the signal, specifically the top-$k$ coordinates for a $k$-sparse signal, which can subsequently be fed to a neural network classifier. In experiments, the classifier is always trained on sparse signals, and at test time, the sparse signal is reconstructed prior to the forward pass. This way, on MNIST and Fashion-MNIST, the algorithm is able to recover large parts of the original accuracy.

https://i.imgur.com/yClXLoo.jpg
Algorithm 1 (see paper for details): The iterative hard thresholding algorithm resulting in provable robustness against $L_0$ attack on images and other signals.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Guo et al. propose to augment black-box adversarial attacks with low-frequency noise to obtain low-frequency adversarial examples as shown in Figure 1. To this end, the boundary attack as well as the NES attack are modified to sample from a low-frequency Gaussian distribution instead from Gaussian noise directly. This is achieved through an inverse discrete cosine transform as detailed in the paper.

https://i.imgur.com/fejvuw7.jpg
Figure 1: Example of a low-frequency adversarial example.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Hosseini and Poovendran propose semantic adversarial examples by randomly manipulating hue and saturation of images. In particular, in an iterative algorithm, hue and saturation are randomly perturbed and projected back to their valid range. If this results in mis-classification the perturbed image is returned as the adversarial example and the algorithm is finished; if not, another iteration is run. The result is shown in Figure 1. As can be seen, the structure of the images is retained while hue and saturation changes, resulting in mis-classified images.

https://i.imgur.com/kFcmlE3.jpg
Figure 1: Examples of the computed semantic adversarial examples.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Karmon et al. propose a gradient-descent based method for obtaining adversarial patch like localized adversarial examples. In particular, after selecting a region of the image to be modified, several iterations of gradient descent are run in order to maximize the probability of the target class and simultaneously minimize the probability in the true class. After each iteration, the perturbation is masked to the patch and projected onto the valid range of [0,1] for images. On ImageNet, the authors show that these adversarial examples are effective against a normal, undefended network.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Li et al. propose camera stickers that when computed adversarially and physically attached to the camera leads to mis-classification. As illustrated in Figure 1, these stickers are realized using circular patches of uniform color. These individual circular stickers are computed in a gradient-descent fashion by optimizing their location, color and radius. The influence of the camera on these stickers is modeled realistically in order to guarantee success.

https://i.imgur.com/xHrqCNy.jpg
Figure 1: Illustration of adversarial stickers on the camera (left) and the effect on the taken photo (right).

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).

Naseer et al. propose to smooth local gradients as defense against adversarial patches. In particular, as illustrated in Figure 1, the local image gradient is computed through convolution. Then, in local, overlapping windows, the gradients are set to zero if the total sum of absolute gradient values exceeds a specific threshold. The remaining gradient map is supposed to indicate regions where it is likely that adversarial patches can be found. Using this gradient map, the image is smoothed, i.e., blurred, afterwards. In experiments, the authors show that this reduces the impact of adversarial patches.

Also find this summary at [davidstutz.de](https://davidstutz.de/category/reading/).