Training Garrabrant inductors to predict counterfactuals discussion post by Tsvi Benson-Tilsen 415 days ago | Abram Demski, Jessica Taylor and Scott Garrabrant like this | discuss The ideas in this post are due to Scott, me, and possibly others. Thanks to Nisan Stiennon for working through the details of an earlier version of this post with me. We will use the notation and definitions given in https://github.com/tsvibt/public-pdfs/blob/master/decision-theory/notation/main.pdf. Let $${{\overline{{\mathbb{P}}}}}$$ be a universal Garrabrant inductor and let $${{\overline{U}}}: {\mathbb{N}}^+ \to {\textrm{Expr}}(2^\omega \to {\mathbb{R}})$$ be a sequence of utility function machines. We will define an agent schema $$({A^{U_{n}}_{n}})$$. We give a schema where each agent selects a single action with no observations. Roughly, $${A^{U_{n}}_{n}}$$ learns how to get what it wants by computing what the $${A^{U_{i}}_{i}}$$ with $$i < n$$ did, and also what various traders predicted would happen, given each action that the $${A^{U_{i}}_{i}}$$ could have taken. The traders are rewarded for predicting what (counterfactually) would be the case in terms of bitstrings, and then their predictions are used to evaluate expected utilities of actions currently under consideration. This requires modifying our UGI and the traders involved to take a possible action as input, so that we get a prediction (a “counterfactual distribution over worlds”) for each action. More precisely, define \begin{aligned} {A^{U_{n}}_{n}} := &\textrm{ let } {\hat{{\mathbb{P}}}}_n := {\textrm{Counterfactuals}}(n)\\ &\;{\textrm{return}}{\operatorname{arg\,max}}_{a \in {\textrm{Act}}} {\hat{{\mathbb{E}}}}_n[a](U_n)\end{aligned} where ${\hat{{\mathbb{E}}}}_n[a](U_n):= \sum_{\sigma \in 2^n} {\hat{{\mathbb{P}}}}_n[a](\sigma) \cdot U_n(\sigma) .$ Here $${\hat{{\mathbb{P}}}}_n$$ is a dictionary of belief states, one for each action, defined by the function $${\textrm{Counterfactuals}}: {\mathbb{N}}^+ \to ({\textrm{Act}}\to \Delta({2^\omega}))$$ using recursion as follows: input: $$n \in {\mathbb{N}}^+$$ output: A dictionary of belief states $${\mathbb{P}}: {\textrm{Act}}\to \Delta({2^\omega})$$ initialize: $${\textrm{hist}}_{n-1} {\leftarrow}$$ array of belief states of length $$n-1$$ for $$i\leq n-1$$: $${\hat{{\mathbb{P}}}}_i {\leftarrow}{\textrm{Counterfactuals}}(i)$$ $$a_i {\leftarrow}{\operatorname{arg\,max}}_{a \in {\textrm{Act}}} \sum_{\sigma \in 2^i} {\hat{{\mathbb{P}}}}_i[a](\sigma) \cdot U_i(\sigma)$$ $${\textrm{hist}}_{n-1}[i] {\leftarrow}{\hat{{\mathbb{P}}}}_i[a_i]$$ for $$(a : {\textrm{Act}})$$: $${\mathbb{P}}[a] {\leftarrow}{\textrm{MarketMaker}}({\textrm{hist}}_{n-1}, {\textrm{TradingFirm}}'(a, a_{\leq n-1}, {\textrm{hist}}_{n-1}))$$ return $${\mathbb{P}}$$ Here, we use a modified form of traders and of the $${\textrm{TradingFirm}}'$$ function from the $$LIA$$ algorithm given in the logical induction paper. In detail, let traders have the type ${\mathbb{N}}^+ \times {\textrm{Act}}\to \textrm{trading strategy}.$ On day $$n$$, traders are passed a possible action $$a \in {\textrm{Act}}$$, which we interpret as “an action that $${A^{U_{n}}_{n}}$$ might take”. Then each trader returns a trading strategy, and those trading strategies are used as usual to construct a belief state $${\mathbb{P}}[a]$$. We pass to $${\textrm{TradingFirm}}'$$ the full history $$a_{\leq n-1}$$ of the actions taken by the previous $${A^{U_{i}}_{i}}$$, since $${\textrm{TradingFirm}}'$$ calls the Budgeter function; that function requires computing the traders’s previous trading strategies, which require passing the $$a_i$$ as arguments. Thus, traders are evaluated based on the predictions they made about logic when given the actual action $$a_n$$ as input. In particular, the sequence $$({\mathbb{P}}_n[a_n])$$ is a UGI over the class of efficient traders given access to the actual actions taken by the agent $${A^{U_{n}}_{n}}$$. This scheme probably suffers from spurious counterfactuals, but feels like a natural baseline proposal.

### NEW DISCUSSION POSTS

This is exactly the sort of
 by Stuart Armstrong on Being legible to other agents by committing to usi... | 0 likes

When considering an embedder
 by Jack Gallagher on Where does ADT Go Wrong? | 0 likes

The differences between this
 by Abram Demski on Policy Selection Solves Most Problems | 0 likes

Looking "at the very
 by Abram Demski on Policy Selection Solves Most Problems | 0 likes

 by Paul Christiano on Policy Selection Solves Most Problems | 1 like

>policy selection converges
 by Stuart Armstrong on Policy Selection Solves Most Problems | 0 likes

Indeed there is some kind of
 by Vadim Kosoy on Catastrophe Mitigation Using DRL | 0 likes

Very nice. I wonder whether
 by Vadim Kosoy on Hyperreal Brouwer | 0 likes

Freezing the reward seems
 by Vadim Kosoy on Resolving human inconsistency in a simple model | 0 likes

Unfortunately, it's not just
 by Vadim Kosoy on Catastrophe Mitigation Using DRL | 0 likes

>We can solve the problem in
 by Wei Dai on The Happy Dance Problem | 1 like

Maybe it's just my browser,
 by Gordon Worley III on Catastrophe Mitigation Using DRL | 2 likes

At present, I think the main
 by Abram Demski on Looking for Recommendations RE UDT vs. bounded com... | 0 likes

In the first round I'm
 by Paul Christiano on Funding opportunity for AI alignment research | 0 likes

Fine with it being shared
 by Paul Christiano on Funding opportunity for AI alignment research | 0 likes