- Model hyperparameters:
- Type of model
- Number of layer
- Size of each layer
- Optimizer hyperparameters:
- Number of children (lambda)
- Number of iterations
- Whether to always update
- Fitness function hyperparameters:
- Optimization goals
- Importance weights for each goal
- Policy hyperparameters:
- Type of policy
- Number of candidates to search
While it's definitely worth tuning the model, optimizer and policy hyperparameters, this doc focuses on understanding and tuning the fitness function weights.
As mentioned in the Mulberry docs, the final fitness of a candidate is a weighted combination of the market level indicators. However, these weights are hyperparameters that must be provided before training and aren't learned in the process. This is because they define what the goal is, adjusting the trade-off between different objectives.
These are the W_i defined in the function above for computing fitness, where S_i is a market level indicator.
As you increase the model complexity or feature set size, you should explore variants using larger values for lambda.
Although we haven't explored it, you may be able to use dresden's hyperparameter search. TODO: I'm also curious about aspects of the modeling approach (like what we were talking about above, intuition for messing with HPs)