Shortcuts

FQF

Overview

FQF was proposed in Fully Parameterized Quantile Function for Distributional Reinforcement Learning. The key difference between FQF and IQN is that FQF additionally introduces the fraction proposal network, a parametric function trained to generate tau in [0, 1], while IQN samples tau from a base distribution, e.g. U([0, 1]).

Quick Facts

  1. FQF is a model-free and value-based distibutional RL algorithm.

  2. FQF only support discrete action spaces.

  3. FQF is an off-policy algorithm.

  4. Usually, FQF use eps-greedy or multinomial sample for exploration.

  5. FQF can be equipped with RNN.

Key Equations or Key Graphs

For any continuous quantile function \(F_{Z}^{-1}\) that is non-decreasing, define the 1-Wasserstein loss of \(F_{Z}^{-1}\) and \(F_{Z}^{-1, \tau}\) by

\[W_{1}(Z, \tau)=\sum_{i=0}^{N-1} \int_{\tau_{i}}^{\tau_{i+1}}\left|F_{Z}^{-1}(\omega)-F_{Z}^{-1}\left(\hat{\tau}_{i}\right)\right| d \omega\]

Note that as \(W_{1}\) is not computed, we can’t directly perform gradient descent for the fraction proposal network. Instead, we assign \(\frac{\partial W_{1}}{\partial \tau_{i}}\) to the optimizer.

\(\frac{\partial W_{1}}{\partial \tau_{i}}\) is given by

\[\frac{\partial W_{1}}{\partial \tau_{i}}=2 F_{Z}^{-1}\left(\tau_{i}\right)-F_{Z}^{-1}\left(\hat{\tau}_{i}\right)-F_{Z}^{-1}\left(\hat{\tau}_{i-1}\right), \forall i \in(0, N).\]

Like implicit quantile networks, a learned quantile tau is encoded into an embedding vector via:

\[\phi_{j}(\tau):=\operatorname{ReLU}\left(\sum_{i=0}^{n-1} \cos (\pi i \tau) w_{i j}+b_{j}\right)\]

Then the quantile embedding is element-wise multiplied by the embedding of the observation of the environment, and the subsequent fully-connected layers map the resulted product vector to the respective quantile value.

The advantage of FQF over IQN can be showed in this picture:

../_images/fqf_iqn_compare.png

Pseudo-code

../_images/FQF.png

Extensions

FQF can be combined with:

  • PER (Prioritized Experience Replay)

    Tip

    Whether PER improves FQF depends on the task and the training strategy.

  • Multi-step TD-loss

  • Double (target) Network

  • RNN

Implementation

Tip

Our benchmark result of FQF uses the same hyper-parameters as DQN except the FQF’s exclusive hyper-parameter, the number of quantiles, which is empirically set as 32. Intuitively, the advantage of trained quantile fractions compared to random ones will be more observable at smaller N. At larger N when both trained quantile fractions and random ones are densely distributed over [0, 1], the differences between FQF and IQN becomes negligible.

The default config of FQF is defined as follows:

class ding.policy.fqf.FQFPolicy(cfg: easydict.EasyDict, model: Optional[torch.nn.modules.module.Module] = None, enable_field: Optional[List[str]] = None)[source]
Overview:

Policy class of FQF algorithm.

Config:

ID

Symbol

Type

Default Value

Description

Other(Shape)

1

type

str

fqf

RL policy register name, refer to
registry POLICY_REGISTRY
this arg is optional,
a placeholder

2

cuda

bool

False

Whether to use cuda for network
this arg can be diff-
erent from modes

3

on_policy

bool

False

Whether the RL algorithm is on-policy
or off-policy

4

priority

bool

True

Whether use priority(PER)
priority sample,
update priority

6

other.eps
.start

float

0.05

Start value for epsilon decay. It’s
small because rainbow use noisy net.

7

other.eps
.end

float

0.05

End value for epsilon decay.

8

discount_
factor

float

0.97, [0.95, 0.999]

Reward’s future discount factor, aka.
gamma
may be 1 when sparse
reward env

9

nstep

int

3, [3, 5]

N-step reward discount sum for target
q_value estimation

10

learn.update
per_collect

int

3

How many updates(iterations) to train
after collector’s one collection. Only
valid in serial training
this args can be vary
from envs. Bigger val
means more off-policy

11

learn.kappa

float

/

Threshold of Huber loss

The network interface FQF used is defined as follows:

class ding.model.template.q_learning.FQF(obs_shape: Union[int, ding.utils.type_helper.SequenceType], action_shape: Union[int, ding.utils.type_helper.SequenceType], encoder_hidden_size_list: ding.utils.type_helper.SequenceType = [128, 128, 64], head_hidden_size: Optional[int] = None, head_layer_num: int = 1, num_quantiles: int = 32, quantile_embedding_size: int = 128, activation: Optional[torch.nn.modules.module.Module] = ReLU(), norm_type: Optional[str] = None)[source]
Overview:

The neural network structure and computation graph of FQF, which combines distributional RL and DQN. You can refer to paper Fully Parameterized Quantile Function for Distributional Reinforcement Learning https://arxiv.org/pdf/1911.02140.pdf for more details.

Interface:

__init__, forward

forward(x: torch.Tensor) Dict[source]
Overview:

Use encoded embedding tensor to predict FQF’s output. Parameter updates with FQF’s MLPs forward setup.

Arguments:
  • x (torch.Tensor):

    The encoded embedding tensor with (B, N=hidden_size).

Returns:
  • outputs (Dict): Dict containing keywords logit (torch.Tensor), q (torch.Tensor), quantiles (torch.Tensor), quantiles_hats (torch.Tensor), q_tau_i (torch.Tensor), entropies (torch.Tensor).

Shapes:
  • x: \((B, N)\), where B is batch size and N is head_hidden_size.

  • logit: \((B, M)\), where M is action_shape.

  • q: \((B, num_quantiles, M)\).

  • quantiles: \((B, num_quantiles + 1)\).

  • quantiles_hats: \((B, num_quantiles)\).

  • q_tau_i: \((B, num_quantiles - 1, M)\).

  • entropies: \((B, 1)\).

Examples:
>>> model = FQF(64, 64) # arguments: 'obs_shape' and 'action_shape'
>>> inputs = torch.randn(4, 64)
>>> outputs = model(inputs)
>>> assert isinstance(outputs, dict)
>>> assert outputs['logit'].shape == torch.Size([4, 64])
>>> # default num_quantiles: int = 32
>>> assert outputs['q'].shape == torch.Size([4, 32, 64])
>>> assert outputs['quantiles'].shape == torch.Size([4, 33])
>>> assert outputs['quantiles_hats'].shape == torch.Size([4, 32])
>>> assert outputs['q_tau_i'].shape == torch.Size([4, 31, 64])
>>> assert outputs['quantiles'].shape == torch.Size([4, 1])

The bellman updates of FQF used is defined in the function fqf_nstep_td_error of ding/rl_utils/td.py.

Benchmark

environment

best mean reward

evaluation results

config link

comparison

Pong

(PongNoFrameskip-v4)

21

../_images/FQF_pong.png

config_link_p

Tianshou(20.7)

Qbert

(QbertNoFrameskip-v4)

23416

../_images/FQF_qbert.png

config_link_q

Tianshou(16172.5)

SpaceInvaders

(SpaceInvadersNoFrame skip-v4)

2727.5

../_images/FQF_spaceinvaders.png

config_link_s

Tianshou(2482)

P.S.:
  1. The above results are obtained by running the same configuration on three different random seeds (0, 1, 2).

References

(FQF) Derek Yang, Li Zhao, Zichuan Lin, Tao Qin, Jiang Bian, Tieyan Liu: “Fully Parameterized Quantile Function for Distributional Reinforcement Learning”, 2019; arXiv:1911.02140. https://arxiv.org/pdf/1911.02140

Other Public Implementations

Read the Docs v: latest
Versions
latest
Downloads
On Read the Docs
Project Home
Builds

Free document hosting provided by Read the Docs.