behavioral cloning

**Behavioral Cloning (BC)** is the **simplest form of imitation learning** — treating the expert's demonstrations as a supervised learning dataset and training a policy to predict the expert's actions from the observed states: $pi(a|s) approx pi_{expert}(a|s)$. **BC Details** - **Dataset**: Expert demonstrations ${(s_i, a_i)}$ — state-action pairs from an expert policy. - **Training**: Supervised learning — minimize $L = sum_i |a_i - pi_ heta(s_i)|^2$ (regression) or cross-entropy (classification). - **Simple**: Just a standard supervised learning problem — any neural network architecture works. - **Distribution Shift**: At test time, small errors compound — the agent visits states not in the training data. **Why It Matters** - **Simplicity**: No reward function, no RL — just supervised learning on demonstrations. - **Compounding Errors**: The main limitation — distributional shift causes errors to accumulate over time. - **Baseline**: BC is the baseline for all imitation learning methods — if BC works well, more complex methods may not be needed. **BC** is **copy the expert** — the simplest imitation learning approach, directly supervised on expert demonstrations.