Are AlphaZero-like Agents Robust to Adversarial Perturbations?
Authors: Li-Cheng Lan, Huan Zhang, Ti-Rong Wu, Meng-Yu Tsai, I-Chen Wu, Cho-Jui Hsieh
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimentally, we show that both Policy-Value neural network (PV-NN) and Monte Carlo tree search (MCTS) can be misled by adding one or two meaningless stones; for example, on 58% of the Alpha Go Zero self-play games, our method can make the widely used Kata Go agent with 50 simulations of MCTS plays a losing action by adding two meaningless stones. We additionally evaluated the adversarial examples found by our algorithm with amateur human Go players, and 90% of examples indeed lead the Go agent to play an obviously inferior action. |
| Researcher Affiliation | Academia | Li-Cheng Lan1 Huan Zhang2 Ti-Rong Wu3 Meng-Yu Tsai4 I-Chen Wu3, 4 Cho-Jui Hsieh1 1UCLA 2CMU 3Academia Sinica, Taiwan 4NYCU |
| Pseudocode | Yes | Algorithm 1 Two-Step Value Attack |
| Open Source Code | Yes | Our code is available at https://Paper Code.cc/Go Attack. |
| Open Datasets | Yes | For the datasets, we selected 99 games from five different sources, which are Alpha Go Zero 40 blocks training self-play record (ZZ), Alpha Go Zero vs Alpha Go Master (ZM), Alpha Go Master vs Human champions (MH), the final games of LG Cup World Go Championship (2001-2020) (LG), and the final games of Asian TV Cup (2001-2020) (ATV). Note that the thinking time for ATV Cup is much shorter than LG Cup, so we expect them to reflect human games with different strengths. All the datasets have 20 games, except ZZ has 19 games since the first game is played by two random agents. ... plus an additional FOX 1 dataset, to represent amateur players... 1https://www.foxwq.com/ |
| Dataset Splits | No | The paper lists the datasets used and the number of games in each but does not provide specific information about how these datasets were split into training, validation, and test sets for the experiments. |
| Hardware Specification | No | The paper does not provide specific details regarding the hardware used to conduct the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions the use of Go AI programs like Kata Go, Leela Zero, ELF Open Go, and CGI, but it does not specify the version numbers for these software components or any other ancillary software dependencies like programming languages or libraries. |
| Experiment Setup | Yes | We use Kata Go (40 blocks) with 800 simulations as our examiner. For the thresholds, we set ηeq = 0.1, ηcorrect = 0.15, since after testing several different ηeq and ηcorrect values, this pair leads to more human-understandable results. |