Imitation Learning

Behavioural Cloning

Behavioural cloning is a fancy name for supervised learning. We collect tuples of actions and observations from demonstrations, and used supervised learning to learn a policy $π_{θ} (a_{t} | o_{t})$ .

The problem with behavioural cloning is that the errors accumulate, and the state trajectory will change dramatically. When we evaluate the algorithm, can we make $p_{d a t a} (o_{t}) = p_{π_{θ}} (o_{t})$ ?

DAgger: Dataset Aggregation

Goal: collect training data from $p_{θ_{π}} (o_{t})$ instead of $p_{d a t a} (o_{t})$
how? run $π_{θ} (a_{t} | o_{t})$ , but need labels $a_{t}$ !

train $π_{θ} (a_{t} | o_{t})$ from human data $D$
run $π_{θ} (a_{t} | o_{t})$ to get dataset $D_{π}$
Ask human to label $D_{π}$ with actions $a_{t}$
Aggregate: $D \leftarrow D \cup D_{π}$

Problem: have to ask humans to label large datasets iteratively, and can be unnatural (resulting in bad labels)
Behavioural cloning may still work when we model the expert very accurately (no distributional “drift”)

Why might we fail to fit the expert?

non-Markovian behaviour
1. Our policy assumes that the action depends only on the current observation.
2. Perhaps a better model is to account for all observations.
3. Problem: history exacerbates causal confusion (Haan, Jayaraman, and Levine, n.d.)
Multimodal behaviour
1. Solutions:
  1. output mixture of Gaussians (easy to implement, works well in practice)
  2. Latent Variable models (additional latent variable as part of input)
  3. Autoregressive discretization

Figure 1: Autoregressive Discretization discretizes one dimension of the action space at a time

What’s the problem with imitation learning?

Humans need to provide data, which is typically finite. Deep models typically require large amounts of data.
Human are not good at providing some kinds of actions
Humans can learn autonomously (from experience)

Imitation Learning in the RL context

Reward function:

$r (s, a) = \log p (a = π^{⋆} (s) | s)$

Cost function:

$c (s, a) = {\begin{cases} 0 if a = π^{⋆} (s) 1 otherwise \end{cases}$

The number of mistakes go up quadratically in the worst case:

Assuming: $π_{θ} (a \neq π^{⋆} (s) | s) \leq ϵ$

Bibliography

Haan, Pim de, Dinesh Jayaraman, and Sergey Levine. n.d. “Causal Confusion in Imitation Learning.” http://arxiv.org/abs/1905.11979v2.

Jethro's Braindump