Graph Neural Networks (GNNs)
Kipf, Thomas N., and Max Welling. "Semi-supervised classification with graph
convolutional networks." arXiv preprint arXiv:1609.02907 (2016).
● Targeting at real-world problems having underlying graph
structures which are non-Euclidean.
○ Euclidean data is data which is sensibly modelled as
being plotted in n-dimensional linear space, e.g.,
images (where the x and y coordinates refer to the
location of each pixel, and the z coordinate refers to its
colour/intensity).
○ Non-Euclidean data: have information of edges,
values associated with nodes, or the directionality of
edges.
Graph Neural Networks (GNNs)
Neural Networks Forward Propagation
●
● w is weights, H is the last hidden states, b is the bias and
sigma is the nonlinear function.
● When i = 0, then H[i]
= X, where X is the input.
Graph Neural Networks
● Sparse.
● Flexible connection.
Graph Neural Networks (GNNs)
Definitions
● Adjacency matrix (A): a N x N matrix filled
with either 0 or 1, where N is the total number
of nodes.
● Node attributes matrix (X): represents the
features or attributes of each node. If there are
N nodes and the size of node attributes is F,
then the shape of this matrix is N x F.
● Edge attributes matrix (E), can also have its
own attributes.
Adjacency matrix
Node attributes matrix (6 x 6)
Graph Neural Networks (GNNs)
Graph
● G = (V, E), denote a graph.
● node feature vectors Xv for v in V
Graph Neural Networks[1]
● Step1: aggregate information from neighbors.
● Step2: combine encoded neighbor’s info with its
own info.
● Step3: use READOUT function to extract the
representation.
Graph Convolutional Neural Network (GCN)
Kipf, Thomas N., and Max Welling. "Semi-supervised classification with graph
convolutional networks." arXiv preprint arXiv:1609.02907 (2016).
Aggregator + Combine:
Representation from the paper
●
GraphSAGE
Hamilton, William L., Rex Ying, and Jure Leskovec. "Inductive representation
learning on large graphs." arXiv preprint arXiv:1706.02216 (2017).
● Aggregator:
● Combine:
Representation from the paper
● Different aggregator
○ Mean, LSTM, Pooling.
GraphSAGE
Hamilton, William L., Rex Ying, and Jure Leskovec. "Inductive
representation learning on large graphs." arXiv preprint arXiv:1706.02216
(2017).
Literature Review
Dong, Jiqian, et al. "A DRL-based Multiagent
Cooperative Control Framework for CAV
Networks: a Graphic Convolution Q Network."
arXiv preprint arXiv:2010.05437 (2020).
● Combining graphic convolution neural
network (GCN) and Deep Q Network (DQN)
to solve the lane changing problem in mixed
traffic with multiple AVs.
● Off-policy Q learning.
Attention (Self-attention)
● Self-attention, also known as intra-attention, is an attention
mechanism relating different positions of a single sequence in
order to compute a representation of the same sequence.
Previous Works
● Iqbal, Shariq, and Fei Sha. "Actor-attention-critic for multi-agent
reinforcement learning." International Conference on Machine Learning.
PMLR, 2019.
Our attention critic is able to
dynamically select which agents to
attend to at each time point during
training, improving performance in
multi-agent domains with complex
interactions.
Graph Convolutional Reinforcement Learning (DGN)
Jiang, Jiechuan, et al. "Graph convolutional
reinforcement learning." arXiv preprint
arXiv:1810.09202 (2018).
● Graph convolution adapts to the dynamics of
the underlying graph of the multi-agent
environment, and relation kernels capture the
interplay between agents by their relation
representations.
● Use self-attention to extract features.
● Off-policy Q learning.
GAT on Autonomous Driving
Wang, Jiawei, et al. "Multi-agent Graph Reinforcement Learning for Connected
Automated Driving", 2020
● On-policy PPO.
● Lane-following.
References
1. Dong, Jiqian, et al. "A DRL-based Multiagent Cooperative Control Framework for CAV Networks: a
Graphic Convolution Q Network." arXiv preprint arXiv:2010.05437 (2020).
2. Gupta, Jayesh K., Maxim Egorov, and Mykel Kochenderfer. "Cooperative multi-agent control using
deep reinforcement learning." International Conference on Autonomous Agents and Multiagent
Systems. Springer, Cham, 2017.
3. Lin, Kaixiang, et al. "Efficient large-scale fleet management via multi-agent deep reinforcement
learning." Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery
& Data Mining. 2018.
4. https://ai.stackexchange.com/questions/11226/what-is-non-euclidean-data#:~:text=Since%20Euclidean
%20spaces%20are%20prototypically,coordinate%20refers%20to%20its%20colour%2F
5. https://towardsdatascience.com/understanding-graph-convolutional-networks-for-node-classification-a
2bfdb7aba7b
6. Kipf, Thomas N., and Max Welling. "Semi-supervised classification with graph convolutional
networks." arXiv preprint arXiv:1609.02907 (2016).