Review

Written during the SERI MATS program under the joint mentorship of John Wentworth, Nicholas Kees, and Janus.

Create a surreal, black and white banner-like image of an agent deciding how to act in an infinitely large lattice-universe with infinitely many people. The agent should be portrayed as thoughtful and contemplative in the center of a vast, lattice-structured universe. This lattice, symbolizing infinity, should be populated with numerous small figures to represent the countless people. The image should have a surreal quality, with abstract patterns and shapes seamlessly integrated into the lattice. The panoramic, banner-style format should enhance the sense of vastness and complexity, emphasizing the enormity of the decision-making process in such an expansive and structured universe.

Preface

In classical game theory, we characterise agents by a utility function and assume that agents choose options which cause maximal utility. This is a pretty good model, but it has some conceptual and empirical limitations which are particularly troublesome for AI safety.

Higher-order game theory (HOGT) is an attempt to rebuild game theory without appealing to either utility functions or maximisation. I think Higher-Order Game Theory is cool so I'm writing a sequence on it.

I'll try to summarise the relevant bits of the literature, present my own minor extensions, and apply HOGT to problems in AI safety

You're reading the first post! Let's get into it.

The role of argmax

For each set , let  be the familiar function which receives a function  and produces the set of element which maximise . A function like  is sometimes called a higher-order function or functional, because it receives another function as input.

Explicitly, .[1]

As you all surely know,  plays a central role in classical game theory. Typically we interpret the set  as the agent's options,[2] and the function  as the agent's task, which assigns a payoff  to each option . We say an option  is optimal to the agent for the task  whenever . Classical game theory is governed by the assumption that agents choose optimal options in whatever task they face, where optimality strictly means utility-maximisation.

Definition 1 (provisional): Let  be any set of options. A task is any function . An option  is optimal for a task  if and only if .

Due to the presence of the powerset operator  in , this model of the agent is possibilistic — for each task , our model says which options are possibly chosen by agent. The model doesn't say which options are probably chosen by the agent — for that we'd need a function . Nor does the model say which options are actually chosen by the agent — for that we'd need a function .[3]

How many options are optimal for the task?Example
Typically there'll be a unique optimal option.
Perhaps multiple options will be optimal.
Perhaps no options are optimal, i.e. every option is strictly dominated by another.
Perhaps every option is optimal, i.e. the task is a constant function.

Exercise 1: Find a set  such that  for every function .


Generalising the functional

The function  is a particular way to turn tasks into sets of options, i.e. it has the type-signature . But there are many functions with the same type-signature (see the table below), so a natural question to ask is... What if we replace  in classical game theory with an arbitrary functional ?

What we get is higher-order game theory.[4] Surprisingly, we can recover many game-theoretic concepts in this more general setting. We can typically recover the original classical concepts from the more general higher-order concepts by restricting our attention to either  or .

So let's revise our definition —

Definition 2 (provisional): Let  be any set of options. An optimiser is any functional . A -task is any function . An option  is -optimal for a task  if and only if .

When clear from context, I'll just say task and optimal.

In higher-order game theory, we model the agents options with a set  and model their task with a function . But (unlike in classical game theory) we're free to model the agent's optimisation with any functional . I hope to persuade you that this additional degree of freedom is actually quite handy.[5]

Higher-order game theory is governed by the central assumption that agents choose -optimal options in whatever -tasks they face, where  is our model of the agent's optimisation. If we observe the agent choosing an option  then that would be consistent with our model, and any observation of a choice  would falsify our model.[6]

Anyway, here is a table of some functionals and their game-theoretic interpretation —

Remarks
This agent will choose an option  which minimises . In classical game theory, this type of optimiser is typically used to model the adversary to the agent modelled by .

This agent will choose an option  which dominates some fixed option . The option  is called the anchor point. It might represent the "default" option, or the "do nothing" option, or the "human-approved" option.

According to Herbert Simon, satisficing is an accurate model of human and institutional decision-making.

Utility-satisficers are a kind of mild optimiser, which might be a safer way to build AI than a full-blown utility-maximiser.

This agent will chooses an option  which maximises the function  up to some fixed slack . Such agents behave like  except that their utility  is measured with finite precision.
This agent will choose an option  which scores better than the average option, given that the option space is equipped with a distribution .

This agent will choose an option in a fixed subset , regardless of the task, e.g. DefectBot and CooperateBot.[7]

Using , we can model a non-agents as a special (degenerate) case of agents.

This agent will choose an option  in the top  quantile, given that the option space is equipped with a distribution .

This is the possibilistic version of Jessica Taylor's quantiliser. Her original probabilistic version is a function , and so wouldn't count as an optimiser according to Definition 2.[8]

This agent will choose an option  which dominates every anchor point in . As special cases, when  we get , and when  is a singleton set we get Simon's satisficer mentioned before. When the set of anchor points is smaller, then the resulting optimiser is less selective, i.e. more options will be optimal.
Exercise 2
Exercise 3This agent will choose an option in the largest equivalence class, where two options are equivalent if they result in the same payoff.

Generalising the payoff space.

Now let's generalise the payoff space to any set , not only . We will think of the elements of  as payoffs in a general sense, relaxing the assumption that the payoffs assigned to the options are real numbers. The function  describes which payoff  would result from the agent choosing the option .

Definition 3 (provisional): Let  be any set of options and  be any set of payoffs. An optimiser is any functional . A -task is any function . An option  is -optimal for a task  if and only if .

This is the final version to this definition today.

This is significantly more expressive! When we are tasked with modelling a game-theoretic situation, we are can pick any set  to model the agent's payoffs![9]

I'll use the notation  to denote the set of functionals , e.g. .

Anyway, here is a table of some functionals and their game-theoretic interpretation —

Payoff spaceRemarks
An optimiser  only needs to be well-defined for bounded utility functions .
An optimiser  only needs to be well-defined for utility functions .

An optimiser  must be well-defined for infinatary utility functions .

For example,  might be the expected total winnings of a gambler employing the gambling strategy . The gambler themselves are modelled by an optimiser . This functional  is characterised by its attitude towards infinite expected profit/loss.

, where  is the distribution monad on .An optimiser  will choose options given a stochastic task . This is the type-signature of risk-averse and risk-seeking maximisers studied in behavioural microeconomics. The field of Portfolio Theory is (largely) the comparison of rival optimisers in .
 is the Levi-Civita Field, an extension of the reals with infinitesimals.

The Levi-Civita field contains infinitesimals like  , as well as infinite values like . In infinite ethics, we encounter tasks , and we can model the infinitary ethicist by an optimiser .

Exercise 4: Solve infinite ethics.

An optimiser  can model different multi-objective optimisers. For example, there's an optimiser which, given multi-objective task , returns those options which are maximal according to the lexicographic ordering, and there's another optimiser which uses the product ordering instead.[10]

Later in this post, we'll encounter an optimiser in  which returns the nash equilibria of -player games, where a task for this optimiser is an -by-payoff matrix .

The field of cooperative bargaining is concerned with different optimisers . A bargaining task  parameterises both the feasibility set  and the disagreement point  where  is the singleton set.

Any preorder 

Suppose that  is any set equipped with a preorder .[11] Then a function  will induce a preorder  on  via .

Let  be the optimiser which chooses the maximal points of , i.e. options which aren't strictly dominated by any other options.

Explicitly 

If  isn't total (i.e. the agent has incomplete preferences over the payoffs) then the resulting optimiser  is less selective (i.e. more options are optimal). In the extreme case, where no options are comparable, then  might choose any option.[12]

Exercise 5: Which optimisers  defined in the previous table can be generalised to any preorder ?

, where  is the option space of the agent.

Occasionally the same set  will serve as both the option space and the payoff space. In this case, a task  represents some transformation of the underlying option space.

There's an optimiser , which choices options which are fixed-points of . That is, . We can use  to model a conservative agent who chooses options which remain untransformed by the task. Note that this optimiser is not consequentialist, because the optimality of an option is not determined by its payoff alone. For example,  but , despite the fact that .


Subjective vs objective optimisers

It's standard practice, when modelling agents and their environments, to use payoff spaces like , etc, but I think this can be misleading.

Consider the following situation —

A robot is choosing an option from a set . There's a function  such that, were the robot to choose the option , then the world would end up in state , where  is something like the set of all configurations of the future light-cone.

You know the robot is maximising over all their options, but you aren't sure what the robot is maximising for exactly — perhaps for paperclips, perhaps for happy humans.

Now, let  be the function which counts the number of paperclips in a light-cone, and let  be the function which counts the number of happy humans.

Here's what classical game theory says about your predicament —

The payoff space is . You know that the robot applies the optimiser , but you don't know whether the robot faces the task  or the task , and hence you don't know whether the robot will choose an option  or .

I call this a subjective account, because the robot's task depends on the robot's preferences. Were the robot to have difference preferences, then they would've faced a different task, and because you don't know the robot's preferences you don't know their task.

However, by exploiting the expressivity of higher-order game theory, we can offer an objective account which rivals the subjective account. In the objective account, the task that the robot faces doesn't depend on the robots preferences —

The payoff space is  itself. You know that the robot faces the task  but you don't know whether the robot applies the optimiser