# Reinforcement Learning

# post


# tools


Reproducible Low-Cost Arm Benchmark, report


# datasets

# workshops confs


RSS 2019

# exploration

Montezuma's revenge and pitfall

Go explore

    1. remember states that havepreviously been visited
    1. first return to a promising state (without exploration),then explore from it
    1. solve simulated environments through exploiting any available means (including by introducing determinism), then robustify (create a policy that can reliably perform the solution) via imitation learning.


# Randomization(exploration, valueF)


# Initialization

for the MuJoCo benchmarks, wider state initialization give you more gains than pretty much any change between RL algorithms and model architectures

# simulation







ODE and Gazebo have the contact support

# browser

three.js stl loader

# testing env

(openai gym)[]

dm control

Arcade Learning Environment(https://github.com/mgbellemare/Arcade-Learning-Environment)


DeepMind Lab


# Model Based


aleatoric (inherent system stochasticity)

epistemic (subjective uncertainty, due to limited data)

Gaussian process

is a stochastic process (a collection of random variables indexed by time or space), such that every finite collection of those random variables has a multivariate normal distribution, i.e. every finite linear combination of them is normally distributed.

Ensembles of bootstrapped models

CEM: samples actions from a distribution closer to previous action samples that yielded high reward

Specifically, aleatoric state variance is the average variance of particles of same bootstrap,

whilst epistemic state variance is the variance of the average of particles of same bootstrap indexes.

# Residual physics

Residual policy learning

# Physics control

fully-actuated in state (q,q˙) at time t if it is able to command any instantaneous acceleration in q underactuated in state (q,q˙) at time t if it is not able to command an arbitrary instantaneous acceleration in q

# State abstraction

semi-mdp(Sutton 1999)


# Distributed

# source of supervision



require env to be inversible, arbitraily resetting env to any states. some env don't.

(Learning to Select and GeneralizeStriking Movements in Robot Table Tennis)[https://www.aaai.org/ocs/index.php/FSS/FSS12/paper/viewFile/5602/5884]


(https://arxiv.org/pdf/1711.00482.pdf)[Learning with latent language]

human preference

# Generalization

Assessing Generalization in Deep Reinforcement Learning

variations in environment dynamics

HalfCheetah under varying joint frictions.

Each environment has three versions: default; random; extreme

EPOpt trains an agent to be robust to environment variations by maximizing a risk-sensitive reward RL2 aims to learn a policy that can adapt to the environment at hand using the observed trajectory

A Dissection of Overfitting and Generalization inContinuous Reinforcement Learning


how to define and diagnose overfitting inMDPs, and how to reduce risks by injecting sufficient training diversity

tasks that received observations from natural images and explore generalization in that setting as well

as soon as there is enough training data diversity in thesimulated environment, deep RL generalizes well

deepRL algorithms show more prominent overfitting when observing natural data.

Results suggest that explicitly learning the dynamics model compounds existing bias in the datain the limited training seed regime

a methodology for detecting overfitting

eval-uation metrics for within-task and out-of-task generalization, consider two mechanisms for injecting noise intothe domain

- an expansion of the initial state distribution, which we implementby applying a multiplier to the initial state chosen.

- Second, we evaluate policy robustness by adding Gaussian noise n ∼ N(0,σ2) directly to theobservation space

training random seeds

generalization error: empirical error difference between test and training


procedural generation of video game levels during training to improve generalization to human-designed levels at test time

safety in grid world


variations in environment dynamics



compare gym and dm environment

  • reward function design
  • does model fit to reward function design?
  • change component behavior
  • hidden instability

# Underactuated Robotics


Quantile Regression Q learning



quantile Q(s,a) to atoms = \sum p_i x_i

sample r, s' from replay buffer, then sample from r + \gamma Z(s', a_\star), quantile atoms to equidistant grids, compute cross-entropy loss

unify cross-entropy and KL divergence:

m is the prob of aligned atoms of $$ r + \gamma Z(x_{x+1}, a^{\star}) $$, and true prob is aligned atoms of Z(x_t,a)

derivatives of KL(m | p_{\theta}) wrt. \theta is derivative of entropy H(m, p_{\theta}), which is the gradient of cross entropy loss function

A Distributional Perspective on Reinforcement Learning

fixed quantile -> variable length gaps

slice to N equal mass, put atoms at median

Huber loss for computing quantile gradient

Wasserstein metric

integrate discrepancy region, different between CDF

W distance is reduced when medians are aligned

why not simple regression: the expectation of the quantiles are not the quantiles of the expectation

tune net

tune the simulation physics from physics in real world

human correction of pose keyframes

correction matrix to transform trajactories

diligent robot hospital service robot

# 3d pose, mesh, kinematic simulation


# sample efficiency

reward shaping behavioral cloning reverse curriculum generation

# others

sticky actions section 5.2

Deep Reinforcement Learning that Matters

motion plan and NN