Tradeoff between desirable properties for baseline choices in impact measures

Impact measures are auxiliary rewards for low impact on the agent’s environment, used to address the problems of side effects and instrumental convergence. A key component of an impact measure is a choice of baseline state: a reference point relative to which impact is measured. Commonly used baselines are the starting state, the initial inaction baseline (the counterfactual where the agent does nothing since the start of the episode) and the stepwise inaction baseline (the counterfactual where the agent does nothing instead of its last action). The stepwise inaction baseline is currently considered the best choice because it does not create the following bad incentives for the agent: interference with environment processes or offsetting its own actions towards the objective. This post will discuss a fundamental problem with the stepwise inaction baseline that stems from a tradeoff between different desirable properties for baseline choices, and some possible alternatives for resolving this tradeoff.

One clearly desirable property for a baseline choice is to effectively penalize high-impact effects, including delayed effects. It is well-known that the simplest form of the stepwise inaction baseline does not effectively capture delayed effects. For example, if the agent drops a vase from a high-rise building, then by the time the vase reaches the ground and breaks, the broken vase will be the default outcome. Thus, in order to penalize delayed effects, the stepwise inaction baseline is usually used in conjunction with inaction rollouts, which predict future outcomes of the inaction policy. Inaction rollouts from the current state and the stepwise baseline state are compared to identify delayed effects of the agent’s actions. In the above example, the current state contains a vase in the air, so in the inaction rollout from the current state the vase will eventually reach the ground and break, while in the inaction rollout from the stepwise baseline state the vase remains intact. 

Possible takeaways from the coronavirus pandemic for slow AI takeoff

As the covid-19 pandemic unfolds, we can draw lessons from it for managing future global risks, such as other pandemics, climate change, and risks from advanced AI. In this post, I will focus on possible implications for AI risk. For a broader treatment of this question, I recommend FLI’s covid-19 page that includes expert interviews on the implications of the pandemic for other types of risks. 

A key element in AI risk scenarios is the speed of takeoff – whether advanced AI is developed gradually or suddenly. Paul Christiano’s post on takeoff speeds defines slow takeoff in terms of the economic impact of AI as follows: “There will be a complete 4 year interval in which world output doubles, before the first 1 year interval in which world output doubles.” It argues that slow AI takeoff is more likely than fast takeoff, but is not necessarily easier to manage, since it poses different challenges, such as large-scale coordination. This post expands on this point by examining some parallels between the coronavirus pandemic and a slow takeoff scenario. The upsides of slow takeoff include the ability to learn from experience, act on warning signs, and reach a timely consensus that there is a serious problem. I would argue that the covid-19 pandemic had these properties, but most of the world’s institutions did not take advantage of them. This suggests that, unless our institutions improve, we should not expect the slow AI takeoff scenario to have a good default outcome. 

2019-20 New Year review

This is an annual post reviewing the last year and making resolutions and predictions for next year. This year’s edition features sleep tracking, intermittent fasting, overcommitment busting, and evaluating calibration for all annual predictions since 2014.

2019 review

AI safety research:

• Wrote an updated version of the relative reachability paper, including an ablation study on design choices.
• Coauthored a paper on modeling AGI Safety frameworks with causal influence diagrams, accepted to IJCAI workshop.
• Wrote a paper on avoiding side effects by considering future tasks, accepted to NeurIPS workshop.
• Co-ran a subteam of the safety team focusing on agent incentive design (ongoing).

AI safety outreach:

• Co-organized FLI’s Beneficial AGI conference in Puerto Rico, a more long-term focused sequel to the original Puerto Rico conference and the Asilomar conference. This year I was the program chair for the technical safety track of the conference.
• Co-organized the ICLR AI safety workshop, Safe Machine Learning: Specification, Robustness and Assurance. This was my first time running a paper reviewing process.
• Gave a talk at the IJCAI AI safety workshop on specification, robustness an assurance problems.
• Took part in the DeepMind podcast episode on AI safety (“I, robot”).

Retrospective on the specification gaming examples list

My post about the specification gaming list was recently nominated for the LessWrong 2018 Review (sort of like a test of time award), which prompted me to write a retrospective (cross-posted here). 

I’ve been pleasantly surprised by how much this resource has caught on in terms of people using it and referring to it (definitely more than I expected when I made it). There were 30 examples on the list when was posted in April 2018, and 20 new examples have been contributed through the form since then.  I think the list has several properties that contributed to wide adoption: it’s fun, standardized, up-to-date, comprehensive, and collaborative.

Some of the appeal is that it’s fun to read about AI cheating at tasks in unexpected ways (I’ve seen a lot of people post on Twitter about their favorite examples from the list). The standardized spreadsheet format seems easier to refer to as well. I think the crowdsourcing aspect is also helpful – this helps keep it current and comprehensive, and people can feel some ownership of the list since can personally contribute to it. My overall takeaway from this is that safety outreach tools are more likely to be impactful if they are fun and easy for people to engage with.

This list had a surprising amount of impact relative to how little work it took me to put it together and maintain it. The hard work of finding and summarizing the examples was done by the people putting together the lists that the master list draws on (Gwern, Lehman, Olsson, Irpan, and others), as well as the people who submit examples through the form. What I do is put them together in a common format and clarify and/or shorten some of the summaries. I also curate the examples to determine whether they fit the definition of specification gaming (as opposed to simply a surprising behavior or solution). Overall, I’ve probably spent around 10 hours so far on creating and maintaining the list, which is not very much. This makes me wonder if there is other low hanging fruit in the safety resources space that we haven’t picked yet. 

I have been using it both as an outreach and research tool. On the outreach side, the resource has been helpful for making the argument that safety problems are hard and need general solutions, by making it salient just in how many ways things could go wrong. When presented with an individual example of specification gaming, people often have a default reaction of “well, you can just close the loophole like this”. It’s easier to see that this approach does not scale when presented with 50 examples of gaming behaviors. Any given loophole can seem obvious in hindsight, but 50 loopholes are much less so. I’ve found this useful for communicating a sense of the difficulty and importance of Goodhart’s Law. 

On the research side, the examples have been helpful for trying to clarify the distinction between reward gaming and tampering problems. Reward gaming happens when the reward function is designed incorrectly (so the agent is gaming the design specification), while reward tampering happens when the reward function is implemented incorrectly or embedded in the environment (and so can be thought of as gaming the implementation specification). The boat race example is reward gaming, since the score function was defined incorrectly, while the Qbert agent finding a bug that makes the platforms blink and gives the agent millions of points is reward tampering. We don’t currently have any real examples of the agent gaining control of the reward channel (probably because the action spaces of present-day agents are too limited), which seems qualitatively different from the numerous examples of agents exploiting implementation bugs. 

I’m curious what people find the list useful for – as a safety outreach tool, a research tool or intuition pump, or something else? I’d also be interested in suggestions for improving the list (formatting, categorizing, etc). Thanks everyone who has contributed to the resource so far!

Classifying specification problems as variants of Goodhart’s Law

(Coauthored with Ramana Kumar and cross-posted from the Alignment Forum.)

There are a few different classifications of safety problems, including the Specification, Robustness and Assurance (SRA) taxonomy and the Goodhart’s Law taxonomy. In SRA, the specification category is about defining the purpose of the system, i.e. specifying its incentives.  Since incentive problems can be seen as manifestations of Goodhart’s Law, we explore how the specification category of the SRA taxonomy maps to the Goodhart taxonomy. The mapping is an attempt to integrate different breakdowns of the safety problem space into a coherent whole. We hope that a consistent classification of current safety problems will help develop solutions that are effective for entire classes of problems, including future problems that have not yet been identified.

The SRA taxonomy defines three different types of specifications of the agent’s objective: ideal (a perfect description of the wishes of the human designer), design (the stated objective of the agent) and revealed (the objective recovered from the agent’s behavior). It then divides specification problems into design problems (e.g. side effects) that correspond to a difference between the ideal and design specifications, and emergent problems (e.g. tampering) that correspond to a difference between the design and revealed specifications.

In the Goodhart taxonomy, there is a variable U* representing the true objective, and a variable U representing the proxy for the objective (e.g. a reward function). The taxonomy identifies four types of Goodhart effects: regressional (maximizing U also selects for the difference between U and U*), extremal (maximizing U takes the agent outside the region where U and U* are correlated), causal (the agent intervenes to maximize U in a way that does not affect U*), and adversarial (the agent has a different goal W and exploits the proxy U to maximize W).

We think there is a correspondence between these taxonomies: design problems are regressional and extremal Goodhart effects, while emergent problems are causal Goodhart effects. The rest of this post will explain and refine this correspondence.

ICLR Safe ML workshop report

This year the ICLR conference hosted topic-based workshops for the first time (as opposed to a single track for workshop papers), and I co-organized the Safe ML workshop. One of the main goals was to bring together near and long term safety research communities.

The workshop was structured according to a taxonomy that incorporates both near and long term safety research into three areas – specification, robustness and assurance.

Specification: define the purpose of the system	Robustness: design system to withstand perturbations	Assurance: monitor and control system activity
Reward hacking Side effects Preference learning Fairness	Adaptation Verification Worst-case robustness Safe exploration	Interpretability Monitoring Privacy Interruptibility

We had an invited talk and a contributed talk in each of the three areas.

2018-19 New Year review

2018 progress

Research / AI safety:

• Wrote a paper on measuring side effects using relative reachability in May, and presented the results at the ICML GoalsRL workshop and the AI safety summer school. Since then, some new approaches have come out using my method as a baseline :).
• Made a list of 30 specification gaming examples in AI (assembled from several existing lists). Since the list was posted in April, 16 new examples have been contributed through the form (thanks everyone!). The list received some attention on Twitter, and I was interviewed about it by Wired and the Times.
• Was in the top 30% of NeurIPS reviewers.
• Gave talks at the Oxford AI Society, EA Global London, etc.
• Got involved in organizing the upcoming ICLR AI safety workshop, Safe Machine Learning: Specification, Robustness and Assurance.

Rationality / effectiveness:

• Attended the CFAR mentoring workshop in Prague, and started running rationality training sessions with Janos at our group house.
• Started using work cycles – focused work blocks (e.g. pomodoros) with built-in reflection prompts. I think this has increased my productivity and focus to some degree. The prompt “how will I get started?” has been surprisingly helpful given its simplicity.
• Stopped eating processed sugar for health reasons at the end of 2017 and have been avoiding it ever since.
  • This has been surprisingly easy, especially compared to my earlier attempts to eat less sugar. I think there are two factors behind this: avoiding sugar made everything taste sweeter (so many things that used to taste good now seem inedibly sweet), and the mindset shift from “this is a luxury that I shouldn’t indulge in” to “this is not food”.
  • Unfortunately, I can’t make any conclusions about the effects on my mood variables because of some issues with my data recording process :(.
• Declining levels of insomnia (excluding jetlag):
  • 22% of nights in the first half of 2017, 16% in the second half of 2017, 16% in the first half of 2018, 10% in the second half of 2018.
  • This is probably an effect of the sleep CBT program I did in 2017, though avoiding sugar might be a factor as well.
• Made some progress on reducing non-research commitments (talks, reviewing, organizing, etc).
  • Set up some systems for this: a spreadsheet to keep track of requests to do things (with 0-3 ratings for workload and 0-2 ratings for regret) and a form to fill out whenever I’m thinking of accepting a commitment.
  • My overall acceptance rate for commitments has gone down a bit from 29% in 2017 to 24% in 2018. The average regret per commitment went down from 0.66 in 2017 to 0.53 in 2018.
  • However, since the number of requests has gone up, I ended up with more things to do overall: 12 commitments with a total of 23 units of workload in 2017 vs 19 commitments with a total of 33 units of workload in 2018. (1 unit of workload ~ 5 hours)

Discussion on the machine learning approach to AI safety

At this year’s EA Global London conference, Jan Leike and I ran a discussion session on the machine learning approach to AI safety. We explored some of the assumptions and considerations that come up as we reflect on different research agendas. Slides for the discussion can be found here.

The discussion focused on two topics. The first topic examined assumptions made by the ML safety approach as a whole, based on the blog post Conceptual issues in AI safety: the paradigmatic gap. The second topic zoomed into specification problems, which both of us work on, and compared our approaches to these problems.

Measuring and avoiding side effects using relative reachability

A major challenge in AI safety is reliably specifying human preferences to AI systems. An incorrect or incomplete specification of the objective can result in undesirable behavior like specification gaming or causing negative side effects. There are various ways to make the notion of a “side effect” more precise – I think of it as a disruption of the agent’s environment that is unnecessary for achieving its objective. For example, if a robot is carrying boxes and bumps into a vase in its path, breaking the vase is a side effect, because the robot could have easily gone around the vase. On the other hand, a cooking robot that’s making an omelette has to break some eggs, so breaking eggs is not a side effect.

(image credits: 1, 2, 3)

How can we measure side effects in a general way that’s not tailored to particular environments or tasks, and incentivize the agent to avoid them? This is the central question of our latest paper.

Specification gaming examples in AI

Update: for a more detailed introduction to specification gaming, check out the DeepMind Safety Research blog post!

Various examples (and lists of examples) of unintended behaviors in AI systems have appeared in recent years. One interesting type of unintended behavior is finding a way to game the specified objective: generating a solution that literally satisfies the stated objective but fails to solve the problem according to the human designer’s intent. This occurs when the objective is poorly specified, and includes reinforcement learning agents hacking the reward function, evolutionary algorithms gaming the fitness function, etc.

While ‘specification gaming’ is a somewhat vague category, it is particularly referring to behaviors that are clearly hacks, not just suboptimal solutions. A classic example is OpenAI’s demo of a reinforcement learning agent in a boat racing game going in circles and repeatedly hitting the same reward targets instead of actually playing the game.

Since such examples are currently scattered across several lists, I have put together a master list of examples collected from the various existing sources. This list is intended to be comprehensive and up-to-date, and serve as a resource for AI safety research and discussion. If you know of any interesting examples of specification gaming that are missing from the list, please submit them through this form.

Thanks to Gwern Branwen, Catherine Olsson, Alex Irpan, and others for collecting and contributing examples!