sac.jl

function BatchSAC(;π::ActorCritic{T, DoubleNetwork{ContinuousNetwork, ContinuousNetwork}}, 
                   S,
                   ΔN=50, 
                   SAC_α::Float32=1f0, 
                   SAC_H_target::Float32 = Float32(-prod(dim(action_space(π)))), 
                   𝒟_train, 
                   SAC_α_opt::NamedTuple=(;), 
                   a_opt::NamedTuple=(;), 
                   c_opt::NamedTuple=(;), 
                   log::NamedTuple=(;), 
                   𝒫::NamedTuple=(;), 
                   param_optimizers=Dict(), 
                   normalize_training_data = true, 
                   kwargs...) where T
    normalize_training_data && (𝒟_train = normalize!(deepcopy(𝒟_train), S, action_space(π)))
    𝒟_train = 𝒟_train |> device(π)
    
    𝒫 = (SAC_log_α = [Base.log(SAC_α)], SAC_H_target = SAC_H_target, 𝒫...)
    BatchSolver(;agent=PolicyParams(π=π, π⁻=deepcopy(π)),
                 S=S,
                 𝒫=𝒫,
                 𝒟_train=𝒟_train,
                 log=LoggerParams(;dir = "log/batch_sac", log...),
                 param_optimizers=Dict(Flux.params(𝒫[:SAC_log_α]) => TrainingParams(;loss=SAC_temp_loss, name="temp_", SAC_α_opt...), param_optimizers...),
                 a_opt=TrainingParams(;loss=SAC_actor_loss, name="actor_", a_opt...),
                 c_opt=TrainingParams(;loss=double_Q_loss(), name="critic_", epochs=ΔN, c_opt...),
                 target_fn=SAC_target(π),
                 kwargs...)
        
end