Deep MCE-IRL#
Reference PDF: papers/econirl_package/primers/deep_mce_irl/deep_mce_irl.pdf.
Deep MCE-IRL uses the maximum causal entropy occupancy-matching objective with a neural reward map. The validated package target is nonlinear reward-map recovery over supplied state encodings and known transitions.
Validation Status#
Pass. The generated artifact comes from
papers/econirl_package/primers/deep_mce_irl/deep_mce_irl_run.py and the
shared known-truth harness.
The primary cell is deep_mce_neural_reward: Shapeshifter with a frozen
nonlinear neural reward, linear supplied features, stochastic known transitions,
32 states, 3 actions, 2,000 individuals, and 80 periods. Action 0 is anchored
at zero reward. It passes 9/9 gates on occupancy residual, normalized
reward/value/Q RMSE, policy TV, and Type A/B/C counterfactual regret.
The finite-theta projection is not the primary success artifact for neural reward truth. Parameter gates are used only when the true reward is linear in a well-conditioned supplied feature matrix. In the neural-feature linear-reward cell, the projection is ill conditioned, so the poor projected-theta cosine is reported as diagnostic rather than claimed as finite-theta recovery.
Usage Scope#
Use this path when transitions are known and the reward can be represented by an MLP over supplied state encodings, with an explicit reward gauge such as an anchor action or absorbing-state anchor. Do not treat raw neural weights as identified structural parameters.
This is not raw FCNN/Objectworld-style spatial reward learning. That would require a convolutional reward network over raw spatial inputs; the current validated implementation learns from supplied encodings.
Artifacts#
PDF source:
papers/econirl_package/primers/deep_mce_irl/deep_mce_irl.texResult generator:
papers/econirl_package/primers/deep_mce_irl/deep_mce_irl_run.pyShared DGP harness:
experiments/known_truth.pyResults:
papers/econirl_package/primers/deep_mce_irl/deep_mce_irl_results.json