FinancialPPOAgentOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for the Financial PPO (Proximal Policy Optimization) trading agent.
Inheritance: ModelOptions → TradingAgentOptions<T> → FinancialPPOAgentOptions<T>
For Beginners
PPO is a robust policy gradient algorithm that prevents large policy updates, leading to stable training. These options extend the base trading agent options with PPO-specific parameters.
How It Works
The defaults are Schulman et al. 2017’s MuJoCo setting (Table 3), the one for separate fully-connected policy and
value networks: a horizon of 2048 steps, Adam at 3e-4, 10 epochs of minibatches of 64, γ = 0.99, λ = 0.95, ε = 0.2
and no entropy bonus. Baselines’ ppo1 (the paper’s code) supplies the rest: Adam’s ε = 1e-5, observations
clipped to ±5 after standardizing, and the value loss at weight 1. CleanRL’s additions are available as options:
MaxGradientNorm = 0.5 and EntropyCoefficient = 0.01 with
ValueCoefficient = 0.5.
Constructors
| Constructor | Summary |
|---|---|
FinancialPPOAgentOptions() |
Creates options with the default values. |
FinancialPPOAgentOptions(FinancialPPOAgentOptions<T> other) |
Creates a copy of another options instance. |
Properties
| Property | Summary |
|---|---|
double AdamEpsilon { get; set; } |
Adam’s ε (1e-5, Baselines’ adam_epsilon). |
double ObservationClipRange { get; set; } |
The bound observations are clipped to after standardizing by their running statistics (5, Baselines’ clip_by_value(obz, -5, 5)). |
double? MaxGradientNorm { get; set; } |
The global gradient-norm bound over both networks, or null for none (the paper’s and Baselines’ setting). |
int MiniBatchSize { get; set; } |
Steps per Adam update (64 in the paper’s MuJoCo setting). |
int NumEpochs { get; set; } |
Optimization epochs over each rollout (K, 10 in the paper’s MuJoCo setting). |
int RolloutLength { get; set; } |
Steps collected between updates (the horizon T, 2048 in the paper’s MuJoCo setting). |