AAAI 2020

• Broad trends: Graph networks, Neuro-symbolics, Hierarchies

Plenary Talks

• Talks available at https://aaai.org/Conferences/AAAI-20/livestreamed-talks/
• Stuart Russel
  • Lesson: Human instructions are uncertain. Minimally invasive preferences. Do not convert everything else into paper clips. Keep "everything else" constant unless asked for.
• Geoffrey Hinton:
  • 3D Transformation must be baked into the neural network. Customize generators. Use set transformers.
  • Read papers: http://akosiorek.github.io/ml/2019/06/23/stacked_capsule_autoencoders.html
  • https://arxiv.org/abs/1810.00825
• Yan Lecun
  • Self supervised learning is going to save the day
• Yoshua Bengio
  • Consciousness is bottleneck principle. Look at Recurrent independent mechanisms
  • https://arxiv.org/abs/1909.10893
• David Cox talk
  • Neuro symbolic concept learning http://nscl.csail.mit.edu/

Interesting papers

• An Intrinsically-­‐Motivated Approach for Learning Highly Exploring and Fast Mixing Policies

  • minimizing || (I - ΠP) . 1 ||₂ is a proxy for maximizing entropy
  • https://arxiv.org/pdf/1907.04662.pdf

• [-] Reinforcement Learningof Risk-­‐Constrained Policies in Markov Decision Processes Tomáš Brázdil, Krishnendu Chatterjee, Petr Novotný, Jiří Vahala

  • https://www.fi.muni.cz/~xnovot18/aaai20.pdf Risk constrained Alpha Go (MCTS). Uses Linear programming to bound the risks on the leaf nodes and propagated them up the monte carlos tree datastructure.
  • https://ieeexplore.ieee.org/abstract/document/6750726
  • http://papers.nips.cc/paper/8032-a-lyapunov-based-approach-to-safe-reinforcement-learning.pdf
  • https://statweb.stanford.edu/~ljanson/papers/Risk_Constrained_Reinforcement_Learning-Chow_ea-2016.pdf

• Few-­‐Shot Bayesian Imitation Learning with Logical Program Policies

  • https://arxiv.org/pdf/1904.06317.pdf
  • symbolic algebra with features

• Off-Policy Evaluation in Partially Observable Environments Guy Tennenholtz, Shie Mannor, Uri Shalit https://arxiv.org/abs/1909.03739

  • " Our work sits at an intersection between the fields of RL and Causal Inference."
  • "In Decoupled POMDPs, observed and unobserved states are separated into two distinct processes, with a coupling between them at each time step."
  • "we demonstrate the use of a well-known approach, Importance Sampling (IS): a reweighting of rewards generated by the be- havior policy, π b , such that they are equivalent to unbiased rewards from an evaluation policy π e ."

• Deep Conservative Policy Iteration

  • Nino Vieillard, Olivier Pietquin, Matthieu Geist
  • https://arxiv.org/abs/1906.09784
  • Another algorithm that offers a different loss function from DQN.

• Deep Model-­‐Based Reinforcement Learning via Estimated Uncertainty
  and Conservative Policy Optimization

  • https://arxiv.org/pdf/1911.12574v1.pdf
  • Estimate uncertainty using ensembles.
  • Upper bound the variance of [Qₛₐʰ]

• Querying to Find a Safe Policy Under Uncertain Safety Constraints in Markov Decision Process Shun Zhang, Edmund H. Durfee, Satinder Singh

  • http://web.eecs.umich.edu/~baveja/Papers/AAAI20-Shun.pdf
  • https://web.eecs.umich.edu/~baveja/Papers/ijcai-2018.pdf
  • "An agent that can be trusted to operate safely should thus only change features the user has explicitly permitted."
  • "It is easier to use linear programming (de Farias and Van Roy 2003) than dynamic programming methods like value itera- tion to find a safely-optimal policy because we can easily add constraints to it."
  • Do not convert everything into paper-clips if asked for a coffee.

• Policy Search by Target Distribution Learning for Continuous Control

  • Chuheng Zhang Yuanqi Li Jian Li
  • https://arxiv.org/abs/1905.11041
  • "It is known that existing policy gradient methods (such as vanilla policy gradient, PPO, A2C) may suffer from overly large gradients when the current policy is close to determin- istic, leading to an unstable training process. We show that such instability can happen even in a very simple environment."

• Tree-­‐Structured Policy based Progressive Reinforcement Learning for Temporally Language Grounding in Video

  • Jie Wu 1 , Guanbin Li 1∗ , Si Liu 2 , Liang Lin
  • http://colalab.org/media/paper/AAAI2020-Tree-Structured.pdf
  • "Inspired by hu- man’s coarse-to-fine decision-making paradigm, we design a tree-structured policy to decompose complex action poli- cies and propose a more reasonable primitive action via two- stages selection, instead of using a flat policy that maps the state feature to action directly (He et al. 2019). As shown in the right half of Figure 2, the tree-structured policy consist- s of a root policy and a leaf policy at each time step. The root policy π r (a rt |s t ) decides which semantic branch will be primarily relied on. The leaf policy π l (a lt |s t , a rt ) consists of five sub-policies, which corresponds to five high-level semantic branches."

• Gradient-­‐Aware Model-­‐based Policy Search

  • Pierluca D'Oro, Alberto Maria Metelli, Andrea Tirinzoni, Matteo Papini, Marcello Restelli
  • https://arxiv.org/abs/1909.04115
  • propagate environment model gradients based on weights based on policy gradients

• Deterministic Value-­‐Policy Gradients

  • https://deepai.org/publication/deterministic-value-policy-gradients
  • Instead of just policy gradients as in DDPG use both Value and policy gradients.

• Safe Linear Stochastic Bandits

  • https://arxiv.org/pdf/1911.09501.pdf
  • "the learner is required to select an arm with an expected reward that is no less than a predetermined (safe) threshold with high probability"
  • P (<Xₜ,θ*> ≥ b) ≥ 1 − δ,

• Planning and Acting with Non-­‐deterministic Events: Navigating between Safe States

• NeoNav: Improving the Generalization of Visual Navigation via Generating Next Expected Observations

  • https://arxiv.org/abs/1906.07207
  • Predict next visual observations using VAEs to improve visual navigation.

• PsyNet: Self-­‐supervised Approach to Object Localization using Point Symmetric Transformation

  • What is a point symmetric transformation
  • https://github.com/FriedRonaldo/PsyNet

• Learning and Reasoning for Robot Sequential Decision Making under Uncertainty

  • https://arxiv.org/abs/1901.05322
  • "In experi- ments, a mobile robot is tasked with estimating human in- tentions using their motion trajectories, declarative contex- tual knowledge, and human-robot interaction (dialog-based and motion-based)"

• Adaptive Trust Region Policy Optimization: Global Convergence and Faster Rates for Regularized MDPs

  • https://arxiv.org/abs/1909.02769
  • Another method using rollout of policy trees for faster convergence.

• Specifying Weight Priors in Bayesian Deep Neural Networks with Empirical Bayes

  • https://arxiv.org/abs/1906.05323
  • Getting weight priors based on mutual information between parameter posterior distribution and predictive distribution.

• Collaborative sampling for Generative Adversarial Networks

  • https://arxiv.org/abs/1902.00813
  • Utilizing discriminator for more informed sampling.