Dealing With Multimodal Distributions

1. Mixture of Gaussians

• Multiple Gaussians, with multiple means and variances (weight for each Guassian)
• Need quite a lot of distributions (exponential in number of parameters)

2. Latent Variable Models

• Ways to solve:
  1. Conditional variational auto encoder
  2. Normalizing flow
  3. Stein variational gradient descent

3. Autoregressive Discretization

• Discrete action space is easy, because we can use softmax
• However, discretizing a continuous action space is pretty difficult
• In autoregressive discretization, we first discretize the first dimension of the action
• Then, sample for the softmax, and have a value for the first dimension
• This feeds into another neural network, which discretizes the second dimension (and repeat)

---

uid: 202008311522 tags: #cs285

---

---

---