# Ω 3

Personal Blog
Crossposted from the AI Alignment Forum. May contain more technical jargon than usual.

author:

• 'Tsvi Benson-Tilsen' title: Training Garrabrant inductors to predict counterfactuals ...

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 be a universal Garrabrant inductor and let be a sequence of utility function machines. We will define an agent schema .

We give a schema where each agent selects a single action with no observations. Roughly, learns how to get what it wants by computing what the with did, and also what various traders predicted would happen, given each action that the 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

where Here is a dictionary of belief states, one for each action, defined by the function using recursion as follows:

<span>input:</span>

<span>output:</span> A dictionary of belief states

<span>initialize: </span> array of belief states of length

<span>for</span> :

<span>for </span> :

<span>return </span>

Here, we use a modified form of traders and of the function from the algorithm given in the logical induction paper. In detail, let traders have the type On day , traders are passed a possible action , which we interpret as “an action that might take”. Then each trader returns a trading strategy, and those trading strategies are used as usual to construct a belief state . We pass to the full history of the actions taken by the previous , since calls the Budgeter function; that function requires computing the traders’s previous trading strategies, which require passing the as arguments.

Thus, traders are evaluated based on the predictions they made about logic when given the actual action as input. In particular, the sequence is a UGI over the class of efficient traders given access to the actual actions taken by the agent .

This scheme probably suffers from spurious counterfactuals, but feels like a natural baseline proposal.

New Comment