Performative Control for Linear Dynamical Systems

Authors: Songfu Cai, Fei Han, Xuanyu Cao

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical results validate our theoretical analysis. and Finally, we conduct experiments on policy-dependent stock investment risk minimization problem. The numerical results validate the effectiveness of our algorithm and theoretical analysis.
Researcher Affiliation Academia Songfu Cai , Fei Han , Xuanyu Cao Department of Electronic and Computer Engineering The Hong Kong University of Science and Technology eesfcai@ust.hk, fhanac@connect.ust.hk, eexcao@ust.hk
Pseudocode Yes Algorithm 1 Repeated Stochastic Gradient Descent (RSGD) Input: Step sizes {ηn, 0 n N}, parameters K, H. Define M = {M : M M} . Initialize M0 M arbitrarily. 1: for n = 0, , N, do 2: Initialize JT = 0. 3: for t = 0, , T 1, do 4: Use control ut = Kxt + Mn [w]H t 1 . 5: Observe At, xt+1; compute noise wt = xt+1 Atxt But. 6: Compute the gradient Mnct (xt, ut) and update JT JT + Mnct (xt, ut) . 7: end for 8: Update Mn+1 Proj M{Mn ηn JT }. 9: end for
Open Source Code No Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: The paper does not provide open access to the data and code.
Open Datasets No In the simulation, we set the number of stocks L = 10 and the number of trading days T = 60. The entries of noise term wt are independently and uniformly drawn from the interval [0, 1]. For each 1 i 10 and each 0 t < 60, we first obtain v(i) t by sampling from a Gaussian distribution, i.e., v(i) t N(log(εt), 0.2), where εt denotes the sensitivity at t-th trading day. The v(i) t is then projected to the interval [ 0.6, 0.6] to obtain v(i) t in (23).
Dataset Splits No In the simulation, we set the number of stocks L = 10 and the number of trading days T = 60. (No explicit mention of training, validation, or test splits. The experiment runs on the simulated data for the full T=60 days).
Hardware Specification No The paper does not provide the type of compute workers, memory, or time of execution. (From NeurIPS Checklist).
Software Dependencies No No specific software dependencies with version numbers are mentioned. The paper describes the algorithm and parameters like 'total number iterations N = 1000, and the stepsize ηn in Algorithm 1 is set to be 0.01'.
Experiment Setup Yes In the simulation, we set the number of stocks L = 10 and the number of trading days T = 60. The initial policy M0 is randomly chosen within the feasible set M = {M : P10 i=1 ml,i = 1, 1 l 10}. The entries of noise term wt are independently and uniformly drawn from the interval [0, 1]. ... The total number iterations N = 1000, and the stepsize ηn in Algorithm 1 is set to be 0.01, 0 n 1000.