Intelligent Agent Foundations Forumsign up / log in
In memoryless Cartesian environments, every UDT policy is a CDT+SIA policy
post by Jessica Taylor 465 days ago | Vadim Kosoy and Abram Demski like this | 1 comment

Summary: I define a memoryless Cartesian environments (which can model many familiar decision problems), note the similarity to memoryless POMDPs, and define a local optimality condition for policies, which can be roughly stated as “the policy is consistent with maximizing expected utility using CDT and subjective probabilities derived from SIA”. I show that this local optimality condition is necesssary but not sufficient for global optimality (UDT).

Memoryless Cartesian environments

I’ll define a memoryless Cartesian environment to consist of:

  1. a set of states \(\mathcal{S}\)
  2. a set of actions \(\mathcal{A}\)
  3. a set of observations \(\mathcal{O}\)
  4. an initial state \(s_1 \in \mathcal{S}\)
  5. a transition function \(t : \mathcal{S} \times \mathcal{A} \rightarrow \Delta \mathcal{S}\), determining the distribution of states resulting from starting in a state and taking a certain action
  6. an observation function \(m : \mathcal{S} \rightarrow \mathcal{O}\), determining what the agent sees in a given state
  7. a set \(\mathcal{S}_T \subset \mathcal{S}\) of terminal states. If the environment reaches a terminal state, the game ends.
  8. a utility function \(U : \mathcal{S}_T \rightarrow [0, 1]\), measuring the value of each terminal state.

On each iteration, the agent observes some observation, and takes some action. Unlike in a POMDP, the agent has no memory of previous observations: the agent’s policy must take into account only the current observation. That is, the policy is of type \(\mathcal{O} \rightarrow \Delta \mathcal{A}\). In this analysis I’ll assume that, for any state and policy, the expected number of iterations in the Cartesian environment starting from that state and using that policy is finite.

Memoryless Cartesian environments can be used to define many familier decision problems (for example, the absent-minded driver problem, Newcomb’s problem with opaque or transparent boxes (assuming Omega runs a copy of the agent to make its prediction), counterfactual mugging (also assuming Omega simulates the agent)). Translating a decision problem to a memoryless Cartesian environment obviously requires making some Cartesian assumptions/decisions, though; in the case of Newcomb’s problem, we have to isolate Omega’s simulation of the agent as a copy of the agent.

Globally and locally optimal policies

Memoryless Cartesian environments are much like memoryless POMDPs, and the following analysis is quite similar to that given in some previous work on memoryless POMDPs: the main difference is that I am targeting (local) optimality given a known world model, while previous work usually targets asymptotic (local) optimality given an unknown world model.

Let us define the expected utility of a particular state, given a policy:

\[V_\pi(s) := U(s) \text{ if } s \in \mathcal{S}_T\] \[V_\pi(s) := \sum_a \pi(a | m(s)) \sum_{s'} t(s' | s, a) V_{\pi}(s') \text{ otherwise}\]

Although this definition is recursive, the recursion is well-founded (since the expected number of iterations starting from any particular state is finite). Note that the agent’s initial expected utility is just \(V_{\pi}(s_1)\). Now we can also define a Q function, determining the expected utility of being in a certain state and taking a certain action:

\[Q_{\pi}(s, a) := \sum_{s'} t(s' | s, a) V_{\pi}(s')\]

Let \(N\) be a random variable indicating the total number of iterations, and \(S_1, ..., S_N\) be random variables indicating the state on each iteration. It is now possible to define the frequency of a given state (i.e. the expected number of times the agent will encounter this state):

\[F_{\pi}(s) := \mathbb{E}\left[\sum_{i=1}^N [S_i = s] \middle| \pi\right]\]

These frequencies are bounded since the expectation of \(N\) is bounded. Given an observation, the agent may be uncertain which state it is in (since multiple states might result in the same observation). It is possible to use SIA to define subjective state probabilities using these frequencies:

\[SIA_{\pi}(s | o) := [m(s) = o] F_{\pi}(s)\]

Note that I’ve defined SIA to return an un-normalized probability distribution; this turns out to be useful later, since it naturally handles the case when the observation \(o\) occurs with probability 0.

How might an agent decide which action to take? Under one approach (UDT), the agent simply computes the globally optimal policy \(\pi\) that results in maximum expected utility (that is, a policy \(\pi\) maximizing \(V_{\pi}(s_1)\)) and takes the action recommended by this policy (perhaps stochastically). While UDT is philosophically satisfying, it is not a very direct algorithm. It would be nice to have a better intuition for how an agent using UDT acts, such that we could (in some cases) derive a polynomial-time algorithm.

So let’s consider a local optimality condition. Intuitively, the condition states that if the agent has a nonzero probability of taking an action \(a\) given observation \(o\), then that action should maximize expected utility (given the agent’s uncertainty about which state it is in). More formally, the local optimality condition states:

\[\forall o \in \mathcal{O}, a \in \mathcal{A}: \pi(a | o) > 0 \rightarrow a \in \arg\max_{a'} \sum_s SIA_{\pi}(s | o) Q_{\pi}(s, a)\]

Philosophically, a policy is locally optimal iff it is consistent with CDT (using SIA probabilities). This local optimality condition is not sufficient for global optimality (for the same reason that not all Nash equilibria in cooperative games are optimal), but it is necessary. The proof follows.

Global optimality implies local optimality

Let \(s\) be a state and \(\pi\) be a policy. Consider a perturbation of the policy \(\pi\): given observation \(o\), the agent will take action \(a_+\) more often, and action \(a_-\) less often. How will this slight perturbation affect expected utility \(V_{\pi}(o)\)?

\[d_{\pi}(o, a_+, a_-, s) := \frac{\partial}{\partial (\pi(a_+ | o) - \pi(a_- | o))} V_{\pi}(o)\]

\[d_{\pi}(o, a_+, a_-, s) = 0 \text { if } s \in \mathcal{S}_T\] \[d_{\pi}(o, a_+, a_-, s) = \sum_{a} \pi(a | m(s)) \sum_{s'} t(s' | s, a) d_{\pi}(o, a_+, a_-, s') + [m(s) = o] (Q_{\pi}(s, a_+) - Q_{\pi}(s, a_-)) \text{ otherwise}\]

This has a natural interpretation: to compute \(d_{\pi}(o, a_+, a_-, s)\), we compute the expected value of simulating a run starting from \(s\) using policy \(\pi\) and summing \(Q_{\pi}(s', a_+) - Q_{\pi}(s', a_-)\) for all visited states \(s'\) with \(m(s') = o\).

To determine the optimal policy, we are concerned with \(d_{\pi}(o, a_+, a_-, s_1)\) for different observations \(o\) and actions \(a_+, a_-\). To compute this, we imagine starting from the state \(s_1\) and following policy \(\pi\), and sum \(Q_{\pi}(s, a_+) - Q_{\pi}(s, a_-)\) for all visited states \(s\) with \(m(s) = o\). This expected sum is actually equivalent to

\[\sum_{s, m(s) = o} F_{\pi}(s) (Q_{\pi}(s, a_+) - Q_{\pi}(s, a_-))\] \[= \sum_{s} SIA_{\pi}(s | o) (Q_{\pi}(s, a_+) - Q_{\pi}(s, a_-))\]

i.e. the expected value of of \(Q_{\pi}(s, a_+) - Q_{\pi}(s, a_-)\) with \(s\) having \(SIA\) probabilities (up to a multiplicative constant). From here the implication should be clear: if a policy \(\pi\) is not locally optimal, then there is some \(o, a_+, a_-\) triple such that a small change in making \(a_+\) more likely and \(a_-\) less likely given observation \(o\) will increase expected utility (just set \(a_-\) to the non-optimal action having nonzero probability given \(o\), and set \(a_+\) to be a better alternative action). So this policy \(\pi\) would not be globally optimal either.


In memoryless Cartesian environments, policies consistent with CDT+SIA are locally optimal in some sense, and all globally optimal (UDT) policies are locally optimal in this sense. Therefore, if we look at (Cartesian) UDT the right way, it’s doing CDT+SIA with some method for making sure the resulting policy is globally optimal rather than just locally optimal. It is not clear how to extend this analysis to non-Cartesian environments where logical updatelessness is important (e.g. agent simulates predictor), but this seems like a useful research avenue.

by Ryan Carey 253 days ago | Jessica Taylor likes this | link

This result features in the paper by Piccione and Rubeinstein that introduced the absent-minded driver problem [1].

Philosophers like decision theories that self-ratify, and this is indeed a powerful self-ratification principle.

This self-ratification principle does however rely on SIA probabilities assuming the current policy. We have shown that conditioning on your current policy, you will want to continue on with your current policy. i.e. the policy will be a Nash Equilibrium. There can be Nash Equilibria for other policies \(\pi'\) however. The UDT policy will by definition equal or beat these from the ex ante point of view. However, others can achieve higher expected utility conditioning on the initial observation i.e. higher \(SIA_{\pi'}(s|o)Q_{\pi'}(s,a)\). This apparent paradox is discussed in [2] [3], and seems to reduce to disagreement over counterfactual mugging.

So why do we like the UDT solution over solutions that are more optimal locally, and that also locally self-ratify? Obviously we want to avoid resorting so circular reasoning (i.e. it gets the best utility ex ante). I think there are some okay reasons:

  1. it is reflectively stable (i.e. will not self-modify, will not hide future evidence) and
  2. it makes sense assuming modal realism or many worlds interpretation (then we deem it parochial to focus on any reference frame other than equal weighting across the whole wavefunction/universe)
  3. it makes sense if we assume that self-location somehow does not
  4. it’s simpler (utility function given weighting 1 across all worlds). In principle, UDT can also include the locally optimal
  5. it transfers better to scenarios without randomization as in Nate + Ben Levenstein’s forthcoming [4].

I imagine there are more good arguments that I don’t yet know.

  1. p19 Piccione, Michele, and Ariel Rubinstein. “On the interpretation of decision problems with imperfect recall.” Games and Economic Behavior 20.1 (1997): 3-24.
  2. Schwarz, Wolfgang. “Lost memories and useless coins: revisiting the absentminded driver.” Synthese 192.9 (2015): 3011-3036.
  4. Cheating Death in Damascus / Nate Soares and Ben Levenstein / Forthcoming






Note that the problem with
by Vadim Kosoy on Open Problems Regarding Counterfactuals: An Introd... | 0 likes

Typos on page 5: *
by Vadim Kosoy on Open Problems Regarding Counterfactuals: An Introd... | 0 likes

Ah, you're right. So gain
by Abram Demski on Smoking Lesion Steelman | 0 likes

> Do you have ideas for how
by Jessica Taylor on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

I think I understand what
by Wei Dai on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

>You don’t have to solve
by Wei Dai on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

Your confusion is because you
by Vadim Kosoy on Delegative Inverse Reinforcement Learning | 0 likes

My confusion is the
by Tom Everitt on Delegative Inverse Reinforcement Learning | 0 likes

> First of all, it seems to
by Abram Demski on Smoking Lesion Steelman | 0 likes

> figure out what my values
by Vladimir Slepnev on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

I agree that selection bias
by Jessica Taylor on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

>It seems quite plausible
by Wei Dai on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

> defending against this type
by Paul Christiano on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

2. I think that we can avoid
by Paul Christiano on Autopoietic systems and difficulty of AGI alignmen... | 0 likes

I hope you stay engaged with
by Wei Dai on Autopoietic systems and difficulty of AGI alignmen... | 0 likes


Privacy & Terms