Adversarial Policies Beat Superhuman Go AIs

Wang, Tony Tong; Gleave, Adam; Belrose, Nora; Tseng, Tom; Miller, Joseph M.; Dennis, Michael D.; Duan, Yingli; Pogrebniak, Viktor; Levine, Sergey; Russell, Stuart D.

doi:10.48550/arxiv.2211.00241

Cited by 5 publications

(4 citation statements)

References 10 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…One might hope that these factors will lead to very reliable performance. However, in [14], we show that these systems still harbor severe weaknesses and can make game-losing mistakes that much weaker humans would not. These mistakes can also be induced by humans without algorithmic assistance, and transfer in a zero-shot setting to different Go systems.…”

Section: Phase 1: Specific Applications and Domainsmentioning

confidence: 82%

Better Bridges Between Model and RealWorld

Pelrine

2023

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

To build better machine learning solutions, we need not only better models but also better bridges through the inputs and outputs to the real world challenges they aim to solve. On the input side, these bridges are the tools one has to work with the available data. On the output side, they are the tools to evaluate and ensure models will generalize and work as intended. This project aims to improve our understanding and resources on both sides, with a particular focus on social good applications. So far, progress has been made in 6 contexts, such as countering misinformation and human trafficking. Future work aims to both build on these specific contexts, as well as leverage their interconnections to produce insights and tools with wide applicability.

show abstract

Section: Phase 1: Specific Applications and Domainsmentioning

confidence: 82%

Better Bridges Between Model and RealWorld

Pelrine

2023

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

show abstract

“…Modern RL methods in multiplayer iterated settings are typically multi-agent [29,30,41]. While our experiments provides initial insights on adversarial collusion, recent studies have also shown that even state of the art multi-agent RL agents can be exploited by adversarial policies [42]. More work would be needed to assess the effectiveness of our attack against algorithms trained using multi-agent RL techniques.…”

Section: Discussionmentioning

confidence: 96%

Adversarial competition and collusion in algorithmic markets

2023

View full text Add to dashboard Cite

Algorithms are now playing a central role in digital marketplaces, setting prices and automatically responding in real time to competitors' behavior. The deployment of automated pricing algorithms is scrutinized by economists and regulatory agencies, concerned about its impact on prices and competition. Existing research has so far been limited to cases where all firms use the same algorithm, suggesting that anti-competitive behaviour might spontaneously arise in that setting. Here, we introduce and study a general anti-competitive mechanism, adversarial collusion, where one firm manipulates other sellers that use their own pricing algorithm. We propose a network-based framework to model the strategies of pricing algorithms on iterated 2 and 3-firm markets. In this framework, an attacker learns to endogenize competitors' algorithms and then derive a strategy to artificially increase its profit at the expense of competitors. Facing a drastic loss of profits, competitors will eventually intervene and revise or turn off their pricing algorithm. To disincentivize this intervention, we show that the attacker can instead unilaterally increase both its profits and the profits of competitors. This leads to a collusive outcome with symmetric and supra-competitive profits, sustainable in the long run. Together, our findings highlight the need for policymakers and regulatory agencies to consider adversarial manipulations of algorithmic pricing, which might currently fall outside of the scope of current competition laws.

show abstract

“…AI-assisted technologies can encourage curiosity, questioning, systematic thinking, trial and error, reasoning, and elaboration (Güss et al 2021). A recently published preprint (Wang et al 2022) shares how curiosity might encourage humans to use AI's 'blind spots.' The underlying argument is that AI-enabled systems encourage workers to reach wide and deep into their knowledge bases, resulting in novel ideas (Althuizen and Reichel 2016).…”

Section: Ai As a General-purpose Tool For Innovative Behaviourmentioning

confidence: 99%

Can artificial intelligence’s limitations drive innovative work behaviour?

Zirar

2023

Rev Manag Sci

View full text Add to dashboard Cite

Artificial intelligence (AI) is deemed to increase workers' productivity by enhancing their creative abilities and acting as a general-purpose tool for innovation. While much is known about AI's ability to create value through innovation, less is known about how AI's limitations drive innovative work behaviour (IWB). With AI's limits in perspective, innovative work behaviour might serve as workarounds to compensate for AI limitations. Therefore, the guiding research question is: How will AI limitations, rather than its apparent transformational strengths, drive workers' innovative work behaviour in a workplace? A search protocol was employed to identify 65 articles based on relevant keywords and article selection criteria using the Scopus database. The thematic analysis suggests several themes: (i) Robots make mistakes, and such mistakes stimulate workers' IWB, (ii) AI triggers 'fear' in workers, and this 'fear' stimulates workers' IWB, (iii) Workers are reskilled and upskilled to compensate for AI limitations, (iv) AI interface stimulates worker engagement, (v) Algorithmic bias requires IWB, and (vi) AI works as a generalpurpose tool for IWB. In contrast to prior reviews, which generally focus on the apparent transformational strengths of AI in the workplace, this review primarily identifies AI limitations before suggesting that the limitations could also drive innovative work behaviour. Propositions are included after each theme to encourage future research.

show abstract

Adversarial Policies Beat Superhuman Go AIs

Cited by 5 publications

References 10 publications

Better Bridges Between Model and RealWorld

Better Bridges Between Model and RealWorld

Adversarial competition and collusion in algorithmic markets

Can artificial intelligence’s limitations drive innovative work behaviour?

Contact Info

Product

Resources

About