struct PgConfig {
int algo;
unsigned long hidden;
unsigned long hiddenLayers;
int act;
float gamma;
float gaeLambda;
float lr;
float entropyCoef;
float valueCoef;
float clipEps;
unsigned long epochs;
unsigned long minibatch;
float clipNorm;
int normalizeAdv;
unsigned long seed;
}Fields
PG_*
A2C / PPO advantages (1 = Monte Carlo, 0 = one-step TD)
PPO ratio clip
PPO passes over a rollout
PPO minibatch size
0 = off
Methods
2 hidden layers of 64 (tanh), gamma 0.99, lambda 0.95, Adam lr 1e-3 (3e-3 for REINFORCE), entropy 0.01, value 0.5, clip 0.2, 4 epochs of 64, grad clip 0.5, normalized advantages.