Safe Haskell	None
Language	Haskell2010

PPO

Description

Proximal Policy Optimization Algorithm

Synopsis

algorithm :: Algorithm
data ActorNetSpec = ActorNetSpec {
- pObsDim :: Int
- pActDim :: Int
}
newtype CriticNetSpec = CriticNetSpec {
- qObsDim :: Int
}
data ActorNet = ActorNet {
- pLayer0 :: Linear
- pLayer1 :: Linear
- pLayer2 :: Linear
}
data CriticNet = CriticNet {
- qLayer0 :: Linear
- qLayer1 :: Linear
- qLayer2 :: Linear
}
data Agent = Agent {
- φ :: ActorNet
- θ :: CriticNet
- logStd :: IndependentTensor
- optim :: Adam
}
mkAgent :: Int -> Int -> IO Agent
saveAgent :: String -> Agent -> IO ()
saveAgent' :: String -> Agent -> IO Agent
loadAgent :: String -> Int -> Int -> Int -> IO Agent
π :: ActorNet -> Tensor -> Tensor
q :: CriticNet -> Tensor -> Tensor
train :: Int -> Int -> HymURL -> TrackingURI -> IO Agent

Documentation

algorithm :: Algorithm Source #

Algorithm ID

data ActorNetSpec Source #

Actor Network Specification

Constructors

ActorNetSpec
Fields pObsDim :: Int pActDim :: Int

Instances

Instances details

Eq ActorNetSpec Source #
Instance details Defined in PPO Methods (==) :: ActorNetSpec -> ActorNetSpec -> Bool # (/=) :: ActorNetSpec -> ActorNetSpec -> Bool #
Show ActorNetSpec Source #
Instance details Defined in PPO Methods showsPrec :: Int -> ActorNetSpec -> ShowS # show :: ActorNetSpec -> String # showList :: [ActorNetSpec] -> ShowS #
Randomizable ActorNetSpec ActorNet Source #	Actor Network Weight initialization
Instance details Defined in PPO Methods sample :: ActorNetSpec -> IO ActorNet #

newtype CriticNetSpec Source #

Critic Network Specification

Constructors

CriticNetSpec
Fields qObsDim :: Int

Instances

Instances details

Eq CriticNetSpec Source #
Instance details Defined in PPO Methods (==) :: CriticNetSpec -> CriticNetSpec -> Bool # (/=) :: CriticNetSpec -> CriticNetSpec -> Bool #
Show CriticNetSpec Source #
Instance details Defined in PPO Methods showsPrec :: Int -> CriticNetSpec -> ShowS # show :: CriticNetSpec -> String # showList :: [CriticNetSpec] -> ShowS #
Randomizable CriticNetSpec CriticNet Source #	Critic Network Weight initialization
Instance details Defined in PPO Methods sample :: CriticNetSpec -> IO CriticNet #

data ActorNet Source #

Actor Network Architecture

Constructors

ActorNet
Fields pLayer0 :: Linear pLayer1 :: Linear pLayer2 :: Linear

Instances

Instances details

Show ActorNet Source #
Instance details Defined in PPO Methods showsPrec :: Int -> ActorNet -> ShowS # show :: ActorNet -> String # showList :: [ActorNet] -> ShowS #
Generic ActorNet Source #
Instance details Defined in PPO Associated Types type Rep ActorNet :: Type -> Type # Methods from :: ActorNet -> Rep ActorNet x # to :: Rep ActorNet x -> ActorNet #
Parameterized ActorNet Source #
Instance details Defined in PPO Methods flattenParameters :: ActorNet -> [Parameter] # _replaceParameters :: ActorNet -> ParamStream ActorNet #
Randomizable ActorNetSpec ActorNet Source #	Actor Network Weight initialization
Instance details Defined in PPO Methods sample :: ActorNetSpec -> IO ActorNet #
type Rep ActorNet Source #
Instance details Defined in PPO type Rep ActorNet = D1 ('MetaData "ActorNet" "PPO" "edelwace-0.1.0.0-NlkYKOSwr4E6F4MQQAmdJ" 'False) (C1 ('MetaCons "ActorNet" 'PrefixI 'True) (S1 ('MetaSel ('Just "pLayer0") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Linear) :: (S1 ('MetaSel ('Just "pLayer1") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Linear) :: S1 ('MetaSel ('Just "pLayer2") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Linear))))

data CriticNet Source #

Critic Network Architecture

Constructors

CriticNet
Fields qLayer0 :: Linear qLayer1 :: Linear qLayer2 :: Linear

Instances

Instances details

Show CriticNet Source #
Instance details Defined in PPO Methods showsPrec :: Int -> CriticNet -> ShowS # show :: CriticNet -> String # showList :: [CriticNet] -> ShowS #
Generic CriticNet Source #
Instance details Defined in PPO Associated Types type Rep CriticNet :: Type -> Type # Methods from :: CriticNet -> Rep CriticNet x # to :: Rep CriticNet x -> CriticNet #
Parameterized CriticNet Source #
Instance details Defined in PPO Methods flattenParameters :: CriticNet -> [Parameter] # _replaceParameters :: CriticNet -> ParamStream CriticNet #
Randomizable CriticNetSpec CriticNet Source #	Critic Network Weight initialization
Instance details Defined in PPO Methods sample :: CriticNetSpec -> IO CriticNet #
type Rep CriticNet Source #
Instance details Defined in PPO type Rep CriticNet = D1 ('MetaData "CriticNet" "PPO" "edelwace-0.1.0.0-NlkYKOSwr4E6F4MQQAmdJ" 'False) (C1 ('MetaCons "CriticNet" 'PrefixI 'True) (S1 ('MetaSel ('Just "qLayer0") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Linear) :: (S1 ('MetaSel ('Just "qLayer1") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Linear) :: S1 ('MetaSel ('Just "qLayer2") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Linear))))

data Agent Source #

PPO Agent

Constructors

Agent
Fields φ :: ActorNet Policy φ θ :: CriticNet Critic θ logStd :: IndependentTensor Standard Deviation (Continuous) optim :: Adam Joint Optimzier

Instances

Instances details

Show Agent Source #
Instance details Defined in PPO Methods showsPrec :: Int -> Agent -> ShowS # show :: Agent -> String # showList :: [Agent] -> ShowS #
Generic Agent Source #
Instance details Defined in PPO Associated Types type Rep Agent :: Type -> Type # Methods from :: Agent -> Rep Agent x # to :: Rep Agent x -> Agent #
type Rep Agent Source #
Instance details Defined in PPO type Rep Agent = D1 ('MetaData "Agent" "PPO" "edelwace-0.1.0.0-NlkYKOSwr4E6F4MQQAmdJ" 'False) (C1 ('MetaCons "Agent" 'PrefixI 'True) ((S1 ('MetaSel ('Just "\966") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 ActorNet) :: S1 ('MetaSel ('Just "\952") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 CriticNet)) :: (S1 ('MetaSel ('Just "logStd") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 IndependentTensor) :*: S1 ('MetaSel ('Just "optim") 'NoSourceUnpackedness 'NoSourceStrictness 'DecidedLazy) (Rec0 Adam))))

mkAgent :: Int -> Int -> IO Agent Source #

Agent constructor

saveAgent :: String -> Agent -> IO () Source #

Save an Agent Checkpoint

saveAgent' :: String -> Agent -> IO Agent Source #

Save an Agent and return the agent

loadAgent :: String -> Int -> Int -> Int -> IO Agent Source #

Load an Agent Checkpoint

π :: ActorNet -> Tensor -> Tensor Source #

Actor Network Forward Pass depending on actionSpace

q :: CriticNet -> Tensor -> Tensor Source #

Critic Network Forward Pass

train :: Int -> Int -> HymURL -> TrackingURI -> IO Agent Source #

Train Proximal Policy Optimization Agent on Environment