Safely controlling the AGI agent reward function

by Koen Holtman5 min read17th Feb 2021No comments

4

CounterfactualsCorrigibilityWireheadingAI
Frontpage

In this fifth post in the sequence, I show the construction a counterfactual planning agent with an input terminal that can be used to iteratively improve the agent's reward function while it runs.

The goal is to construct an agent which has has no direct incentive to manipulate this improvement process, leaving the humans in control.

The reward function input terminal

I will define an agent with an input terminal can be used to improve the reward function of an agent. The terminal contains the current version of the reward function, and continuously sends it to the agent's compute core::

This setup is motivated by the observation that it is unlikely that fallible humans will get a non-trivial AGI agent reward function right on the first try, when they first start it up. By using the input terminal, they can fix mistakes, while the agent keeps on running, if and when such mistakes are discovered by observing the agent's behavior.

As a simplified example, say that the owners of the agent want it to maximize human happiness, but they can find no way of directly encoding the somewhat nebulous concept of human happiness into a reward function. Instead, they start up the agent with a first reward function that just counts the number of smiling humans in the world. When the agent discovers and exploits a first obvious loophole in this definition of happiness, the owners use the input terminal to update the reward function, so that it only counts smiling humans who are not on smile-inducing drugs.

Unless special measures are taken, the addition of an input terminal also creates new dangers. I will illustrate this point by showing the construction of a dangerous agent ITF further below.

Design and interpretation of the learning world

As a first step in defining any agent with an input terminal, I have to define a model of a learning world which has both the agent and its the input terminal inside it. I call this world the learning world, because the agent in it is set up to learn the dynamics of its learning world environment.

See this earlier post in the sequence for a general introduction to the graphical language I am using to define world models and agents.

As a first step to constructing the learning world diagram, I take the basic diagram of an agent interacting with its environment:

To model the input terminal, I then split each environment state node into two components:

The nodes represent the signal from the input terminal, the subsequent readings by the agent's compute core of the signal which encodes a reward function, and the nodes model all the rest of the agent environment state.

I then add the observational record keeping needed to inform online machine learning. I add two separate time series of observational records: and . The result is the learning world diagram :.

In the case that the learning world is our real world, the real input terminal will have to be built using real world atoms (and freely moving subatomic particles).

I use the modeling convention that the random variables represent only the observable digital input terminal signal as received by the agent's compute core. The atoms that make up the input terminal are not in , they are part of the environment state modeled in the variables.

Unsafe factual planning agent ITF

I will now draw a 'standard' factual planning world that models the full mechanics of the learning world, define the ITF agent with it, and show why this agent is unsafe.

ITF: The factual input terminal agent has the learning world where , with defined by the factual planning world , where , , and .

The planning world reward function uses a form of indirect referencing: it applies the function as read from the input terminal in the current time step to compute the reward for that time step.

As I discussed in the previous post, indirection is a powerful tool, but it also often introduces new risks to the setup, typically risks of reward tampering or wireheading.

The ITF planning world agent can massively improve its planning world utility by using a policy where it walks to the input terminal to input the new reward function . Clearly, this is not what we want, if the optimal planning world policy is to input , then the real world (learning world) agent, which copies the actions determined by the planning world policy, would end up inputting this function too.

If some people in the real world want to block the ITF agent from changing its own reward function, then the agent might end up harming these people. It will harm them even if its initial planning world reward function has a penalty term that subtracts from the planning world utility for every human harmed. The agent does not have to come up with any smart way to harm people without triggering this penalty term: it will gladly harm all 500 people guarding the input terminal because is much a larger number than .

Control incentives, indifference, and paths to value

An important feature of causal influence diagram based agent models is that they allow us to generalize from the above thought experiment.

Causal influence diagrams offer a mathematically well-defined way to talk about control incentives and indifference, where indifference is the lack of a control incentive. (More mathematical details are in section 9 of my paper. The mathematical relation between incentives and causal influence diagrams is also explored here.)

The general problem is that the ITF planing world agent has an incentive to control the future signal coming from the input terminal. This incentive is caused by the presence of paths to value in its planning world diagram , for example by the path from to via and to the utility node .

When we want to maximize the probability that the humans will stay in control of the terminal, we have to suppress the agent's incentive to control.

The general technique is discussed in section 9 of the paper: we can make the agent indifferent about what happens to the future input terminal signal by removing all the offending paths to value.

Safer counterfactual planning agent ITC

To construct the safer ITC agent, I remove the offending paths to value by taking the planning world diagram :

and re-routing many of the arrows in it. I end up with the counterfactual planning world diagram :

The rerouted arrows are drawn in green. I also deleted the arrows that go from the nodes to the nodes.

I then define the safer ITC agent by using this counterfactual planning world:

ITC: The counterfactual input terminal agent has the learning world where , where is defined by the planning world , where , , and .

If the initial reward function in the ITC agent planning world rewards paperclip production, then the planning world ITC agent lacks any direct incentive to perform actions that take away resources from paperclip production in order to influence what happens to its input terminal signal, as modeled by the nodes . This is because in the planning world, the state of these nodes has absolutely no influence, either positive or negative, on how the values in the agent's yellow utility nodes are computed.

Corrigibility of the ITC agent

As the learning world ITC agent we defined takes actions defined by the planning world policies, the ITC agent is also indifferent about the value of its future terminal signal. It also has no direct incentive to manipulate the human decision processes that will determine the direction of these updates. As explained in the previous post, this is a form of corrigibility.

But the ITC agent is not perfectly corrigible under all conditions, as it might develop indirect incentives to block the use of the input terminal. For a detailed example, see the discussion about stop button failure modes in this earlier post.

Simplifying the ITC agent design

We can simplify the ITC agent design by removing the nodes from the planning world, as these have no effect on the reward we want the agent to calculate. This in turn allows us to remove the observational records and machine learning function needed to compute the planning world parameter .

These steps produce the following diagrams and agent definition:

CITC: The compact version of the counterfactual input terminal agent has the learning world where , where is defined by the planning world , where and .

Relation between counterfactual planning and indifference methods

In my 2020 paper here, I used non-graphical MDP models and indifference methods to define a similar safe agent with an input terminal, called the agent. When used as a specification for a real-life agent compute core, the agent definition in that paper produces exactly the same corrigible agent compute core behavior as the ITC agent definition above.

The main difference is that the indifference methods based construction of in that paper is more opaque than the counterfactual planning based construction of ITC above.

The agent is constructed by including a complex balancing term in its reward function, were this term can be interpreted as occasionally creating extra virtual worlds inside the agent's compute core. Counterfactual planning constructs a different set of virtual worlds called planning worlds, and these are much easier to interpret.

My 2020 paper includes some dense mathematical proofs to show that the agent has certain safety properties. Counterfactual planning offers a vantage point which makes the same safety properties directly visible in the ITC agent construction, via a lack of certain paths to value in the planning world.

So overall, my claim is that counterfactual planning offers a more general and transparent way to achieve the corrigibility effects that can be constructed via balancing term based indifference methods.

Simulations of ITC agent behavior

See sections 4, 6, 11, and 12 of my 2020 paper for a more detailed discussion of the behavior of the agent, which also applies to the behavior of the ITC agent. These sections also show some illustrative agent simulations.

Section 6 has simulations where the agent will develop, under certain conditions, an indirect incentive causing it to be less corrigible. Somewhat counter-intuitively, that incentive gets fully suppressed when the agent gets more powerful, for example by becoming more intelligent.

4

New Comment