Safe Haskell	None
Language	Haskell2010

RPB.HER

Description

Hindsight Experience Replay

Synopsis

data Strategy
- = Final
- | Random
- | Episode
- | Future
data Buffer a = Buffer {
- states :: !a
- actions :: !a
- rewards :: !a
- states' :: !a
- dones :: !a
- targets :: !a
- targets' :: !a
}
mkBuffer :: Buffer Tensor
empty :: Buffer Tensor
size :: Buffer Tensor -> Int
push :: Int -> Tensor -> Buffer Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Buffer Tensor
push' :: Int -> Buffer Tensor -> Buffer Tensor -> Buffer Tensor
push'' :: Int -> Buffer Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Buffer Tensor
drop' :: Buffer Tensor -> Buffer Tensor
drop :: Int -> Buffer Tensor -> Buffer Tensor
envSplit :: Int -> Buffer Tensor -> [Buffer Tensor]
epsSplit :: Buffer Tensor -> [Buffer Tensor]
sample :: Tensor -> Buffer Tensor -> Buffer Tensor
sampleTargets :: Strategy -> Int -> Tensor -> Buffer Tensor -> IO (Buffer Tensor)
asRPB :: Buffer Tensor -> Buffer Tensor
targetCriterion :: Map String Bool -> Tensor

Documentation

data Strategy Source #

Hindsight Experience Replay Strategies for choosing Goals

Constructors

Final	Only Final States are additional targets
Random	Replay with `k` random states encountered so far (basically RPB)
Episode	Replay with `k` random states from same episode.
Future	Replay with `k` random states from same episode, that were observed after

Instances

Instances details

Eq Strategy Source #
Instance details Defined in RPB.HER Methods (==) :: Strategy -> Strategy -> Bool # (/=) :: Strategy -> Strategy -> Bool #
Show Strategy Source #
Instance details Defined in RPB.HER Methods showsPrec :: Int -> Strategy -> ShowS # show :: Strategy -> String # showList :: [Strategy] -> ShowS #

data Buffer a Source #

Strict Simple/Naive Replay Buffer

Constructors

Buffer
Fields states :: !a States actions :: !a Actions rewards :: !a Rewards states' :: !a Next States dones :: !a Terminal Mask targets :: !a Actual Episode Target targets' :: !a Augmented Target

Instances

Instances details

Functor Buffer Source #
Instance details Defined in RPB.HER Methods fmap :: (a -> b) -> Buffer a -> Buffer b # (<$) :: a -> Buffer b -> Buffer a #
Eq a => Eq (Buffer a) Source #
Instance details Defined in RPB.HER Methods (==) :: Buffer a -> Buffer a -> Bool # (/=) :: Buffer a -> Buffer a -> Bool #
Show a => Show (Buffer a) Source #
Instance details Defined in RPB.HER Methods showsPrec :: Int -> Buffer a -> ShowS # show :: Buffer a -> String # showList :: [Buffer a] -> ShowS #

mkBuffer :: Buffer Tensor Source #

Create a new, empty HER Buffer on the GPU

empty :: Buffer Tensor Source #

Create an empty HER Buffer

size :: Buffer Tensor -> Int Source #

How many Trajectories are currently stored in memory

push :: Int -> Tensor -> Buffer Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Buffer Tensor Source #

Calculate reward and done and Push new memories into Buffer

push' :: Int -> Buffer Tensor -> Buffer Tensor -> Buffer Tensor Source #

Push one buffer into another one

push'' :: Int -> Buffer Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Tensor -> Buffer Tensor Source #

Alternative Push if tensors are not in a buffer yet

drop' :: Buffer Tensor -> Buffer Tensor Source #

Drop everything after last done (used for single episode)

drop :: Int -> Buffer Tensor -> Buffer Tensor Source #

Drop number of entries from the beginning of the Buffer

envSplit :: Int -> Buffer Tensor -> [Buffer Tensor] Source #

Split buffer collected from pool by env

epsSplit :: Buffer Tensor -> [Buffer Tensor] Source #

Split a buffer into episodes, dropping the last unfinished

sample :: Tensor -> Buffer Tensor -> Buffer Tensor Source #

Get the given indices from Buffer

sampleTargets :: Strategy -> Int -> Tensor -> Buffer Tensor -> IO (Buffer Tensor) Source #

Sample Additional Goals according to Strategy (drop first). Random is basically the same as Episode you just have to give it the entire buffer, not just the episode.

asRPB :: Buffer Tensor -> Buffer Tensor Source #

Convert HER Buffer to RPB for training

targetCriterion :: Map String Bool -> Tensor Source #

Convert target predicate map to boolean mask tensor