🌟 What is this article about?

In order to use cram_bandit(), users must supply a matrix of action selection probabilities $\pi_t(X_j, A_j)$ for each combination of policy update $t$ and context $j$ in the historical dataset.

While some environments log these probabilities directly, many contextual bandit libraries (such as contextual) only store policy parameters (e.g., regression coefficients) without explicit probability tracking.

This article explains how Cram Bandit Helpers reconstruct $\pi_t(X_j, A_j)$ from these parameters for common policies:

Both theoretical formulas and practical code snippets are provided.

🛠️Policy parameters explained

When using linear bandit algorithms like Epsilon-Greedy, LinUCB, or Thompson Sampling, each arm $k$ maintains summary statistics (parameters) to estimate the expected reward:

• $A_k$ is the Gram matrix:
  $$A_k = X_k^T X_k$$ where $X_k$ is the matrix of feature vectors (contexts) for all rounds where arm $k$ was selected.
  ➔ Interpretation: $A_k$ captures the amount of information (and correlation structure) about the features for arm $k$. It plays the role of a “feature covariance matrix.”

• $b_k$ is the response vector:
  $$b_k = X_k^T y_k$$ where $y_k$ are the observed rewards for arm $k$.
  ➔ Interpretation: $b_k$ captures the relationship between the observed rewards and the contexts for arm $k$.

These sufficient statistics allow the policy to compute the Least Squares estimate for the reward model:

$$\theta_k = A_k^{-1} b_k$$

where:

• $\theta_k$ is the estimated coefficient vector that predicts the expected reward of arm $k$ as a function of the context.

Thus:

• $A_k$ tells us how confident we are about $\theta_k$ (smaller eigenvalues of $A_k$ imply more uncertainty).
• $b_k$ provides the observed signal (reward-weighted context features) to fit $\theta_k$.

The policy selects an action based on the $\theta_k$ of each arm $k$ and then observe the reward associated with this choice, which is used to update the parameters $A_k$ and $b_k$ of the policy.

✨ Epsilon-Greedy Policy

get_proba_c_eps_greedy(eps, A_list, b_list, contexts, chosen_arms)

🔢 LinUCB Disjoint Policy with $\varepsilon$-Greedy

get_proba_ucb_disjoint(alpha, eps, A_list, b_list, contexts, chosen_arms)

🤓 Thompson Sampling (LinTS)

get_proba_thompson(sigma, A_list, b_list, contexts, chosen_arms)

👨‍💻 Practical Workflow

When using your bandit policy in practice:

1. Record action choices, contexts, and policy parameters (e.g., $A$, $b$)
2. Calculate the action selection probabilities. If your policy is within the ones presented above, please feel free to rely on our helper functions to build $\pi$.
3. Feed pi, arm, and reward into cram_bandit() for evaluation of your policy.

🧪 Estimand Calculation in cram_bandit_sim()

The following only concerns the simulation framework we implemented for benchmarking purposes.

Once the policies are reconstructed, we compute their true expected value — referred to as the estimand — by applying the learned policy to independent contexts and evaluating it against the known reward function used in the simulation.

This is done via:

compute_estimand(data_group, list_betas, policy, policy_name, batch_size, bandit)

Accurately computing the estimand is critical for properly assessing the bias and confidence interval coverage of the Cram estimate in our simulations.

📂 Useful Links

• contextual package: original framework
• cram_bandit(): Cram evaluation for contextual bandits
• cram_bandit_sim(): Full simulation engine with automatic pi estimation

🌟 Acknowledgments

These helper functions were designed to faithfully reconstruct action probabilities for the policies implemented in contextual, while enabling reproducible Cram-based evaluation.