# 4

The following is an edited excerpt from the Preliminaries and Background sections of my now completed MSc thesis in Artificial Intelligence from the University of Amsterdam.

In the thesis, we set out to tackle the issue of Goal Misgeneralization (GMG) in Sequential Decision Making (SDM)[1] by focusing on improving task specification. Below, we first link GMG to causal confusion, motivating our approach. We then outline specifically what we mean by task specification, and later discuss the implications for our own definition of GMG.

I am sharing this here because I believe the ideas presented here are at least somewhat interesting and have not seen them discussed elsewhere. We ultimately did not publish the thesis, so rather than keeping these to myself, I figured I'd at least share them here.

You can find the full thesis along with its code here.

## Causal Confusion and Goal Misgeneralization

Inspired by the works of Gupta et al. (2022) and Kirk and Krueger (2022), we hold the view that GMG is a direct consequence of causal confusion (CC) (de Haan, Jayaraman, and Levine 2019). This is the phenomenon by which a learner incorrectly identifies the causal model underlying its observations and/or behaviour. This is typically due to spurious correlations between the true cause for a random event and some other variable that does not causally model . We posit that CC may lead to GMG when the confounding variable, i.e. the variable spuriously correlated with the causal factor, is easier to learn.

Accordingly, we note that GMG may therefore be addressed by tackling CC itself. In light of this, we can distinguish three approaches. The first involves performing causal inference with the assistance of interventions on the data so to better discover the underlying causal model. This is the main approach of de Haan, Jayaraman, and Levine (2019). The second approach simply increases the variability of the training data so as to reduce the likelihood of spurious correlations. This is the main approach of Langosco et al. (2022). The final approach focuses on improving the expressiveness of the task specification. We hypothesize that overly coarse specifications may lead to ambiguity in which task is being requested, increasing the chance of causal confusion. We provide more detail in the next sections.

While each of these approaches have merit, we decide to focus on the third. Our motivation is manifold. First, we expect implementations under the first approach to become increasingly more difficult as the field shifts towards offline-learning (Lange, Gabel, and Riedmiller 2012; Levine et al. 2020; Prudencio, Maximo, and Colombini 2022). Secondly, while the simplicity of the second approach coupled with recent advancements in scaling laws (Kaplan et al. 2020; Hoffmann et al. 2022) is promising, we note that increasing the variability of the training data has no guarantee of de-correlating confounding variables, especially when the spurious correlations are unknown, rendering estimating how much and what kind of variability to work on potentially difficult for more insidious cases of GMG (Kirk and Krueger 2022). We choose to focus on the approach of improving task specification not only because we view it as an under-explored option, but more importantly because, as we will outline below, we view GMG as intrinsically tied to multi-task learning (Caruana 1997), which itself is intrinsically tied to task specification.

Task specification is the scenario in which a requester specifies a task to be performed by an actor [2]. In SDM, The requester expresses a high-level representation of the ideal trajectory of state-action pairs, corresponding to the task they would like to be performed. We specifically allow high-level representations of trajectories because it can occur that the requester does not know exactly what sequence of state-action pairs they want, and are typically more interested in more abstract, higher level desiderata anyway.

The actor is necessarily a multi-task policy, as otherwise task-specification would be futile. The actor receives and "interprets" it by using it as a conditional variable on its policy. Like M. Cho, Jung, and Sung (2022), we therefore write the actor's policy as , where represents an encoding of the intended task. We underline that can in principle take any form and originate from any source. Examples include rewards, one-hot encodings, demonstrations, preferences (Christiano et al. 2017), formal language (Bansal 2022), natural language, et cetera.

## Specification and causal confusion

Suppose we have some latent notion , an abstraction encapsulating some semantic information, that we wish to communicate. The notion is latent, i.e. not observed directly, and we can instead communicate it through some language which maps the notion to some corresponding expression . Note that there can be more than one corresponding expression per notion. In general, the mapping between notion and language expression is many-to-many. Under our task specification framework from above, the task we wish to specify is the notion we wish to communicate , and the high-level representation is the expression we use to communicate it.

In the context of communication, a notion and its corresponding expressions , can be treated as random variables. This assumption can be made given the wide, almost infinite range of possible notions one may wish to communicate, and similarly to the wide range of ways in which a notion can be expressed. These lead to uncertainty which we can treat probabilistically with random variables.

We can therefore quantify the information content of a given notion or expression using the concept of entropy (Shannon 1948). Entropy effectively quantifies the average level of uncertainty or "surprise" associated with a random variable. For a discrete random variable , its entropy is defined as

where is the probability mass function of , and the summation is over all possible outcomes of . A higher entropy indicates greater uncertainty and thus greater information content. If an outcome is highly uncertain, it means we have very little prior knowledge about what that outcome will be. Therefore, learning the actual outcome provides us with a substantial amount of new information. Conversely, if an event is certain to occur, then learning that this event has indeed occurred doesn't provide us with any new information because we already knew it would happen. Thus, a higher entropy indicates greater uncertainty and thus greater information content.

The entropy of a given notion and an expression of it therefore serves as the measure of their respective information content. For a notion, we can write

where is the probability of notion being the one intended for communication. For an expression, we can write

where is the probability of expression being the one used for communication.

will typically be a compressed representation of . In other words, the mapping between notion and expression is not necessarily lossless in terms of information

This compression can be either intrinsic or extrinsic. The former case corresponds to compression that occurs due to the fundamentally limited expressivity of the language . For example, a language that lacks the grammar and/or vocabulary for expressing negation, will be fundamentally limited from expressing the notion of absence.

Extrinsic compression is compression that occurs due to reasons external to the language itself. This is typically the communicator choosing to use a coarser expression of the notion. For example, choosing to communicate "go to the block" rather than "breathe in, activate your muscles such that your right thigh lifts your right foot off the ground and forward, breathe out, breathe in, ...".

Compression, whether intrinsic, extrinsic or either, can lead to ambiguity. These are cases where the same expression , due to underspecification, maps to multiple semantically different notions . We view this as a potential avenue for causal confusion to occur.

For instance, under our definitions, we can frame rewards as a language used to communicate some notion of a desired task to SDM agents. When our rewards are underspecified, they can over-compress our task notion, such that the same reward maps to multiple tasks. The policy may therefore suffer from causal confusion and learn to pursue the wrong task.

We therefore posit that causal confusion and hence GMG can be addressed by focusing on how we specify the task, so to reduce ambiguity in the task specification. We move away from rewards (Vamplew et al. 2022) and instead leverage the potentially much higher expressiveness of natural language, spurred by recent advancements in the field of natural language processing (NLP) (Devlin et al. 2019; Brown et al. 2020; Touvron et al. 2023). For a given notion , assuming the same amount of engineering effort, we expect the compression faced by the language of rewards to be higher than the compression faced by natural language , i.e. we expect the following

We reason that the language of rewards faces higher intrinsic compression due to its scalar nature, rendering it more difficult to capture nuance than what would be possible with the multidimensionality and compositionality of natural language, which could not only encode more information directly, but could also allow for factored representations which may more easily be leveraged for generalization. Similarly, we expect the language of rewards to also face higher extrinsic compression when compared to natural language. We reason that task specification is a communication problem, and to this end natural language is the most natural or "comfortable" interface we have as communicators. Rewards, while succinct, may at times be awkward to specify due to the nature of the tasks. This is for instance the case for sparse rewards awarded only upon task completion, or for the denser proxy rewards awarded in the process of reward shaping (Ng, Harada, and Russell 1999).

## Defining GMG in the context of multi-task learning

Goal Misgeneralization is inherently Multi-task. Indeed, all definitions and examples of GMG so far have implicitly defined a multi-task setup, with some goal task and some other confounding task . After all, the definition of GMG implies the existence of at least one other task beyond the one intended by the designers, as without such a task, it would be impossible for the model to pursue it. We instead choose to explicitly define this multi-task setup, relying on the framework from Wilson et al. (2007).

Specifically, let be a set of discrete episodic tasks. This could for example the set of all tasks with natural language instructions , following the notion and expression notation from the previous section. Let and be the distributions from which the tasks are sampled during training and testing respectively. Each task then defines a separate MDP , such that the reward and transition functions differ by task. At training time we try to find a task-conditioned policy

with an objective conductive to good performance across the tasks. For multi-task RL, such an objective maximizes the expected reward over the distribution of tasks, i.e.

where is the horizon of time steps and is the discount factor. For multi-task IL, such an objective minimizes the expected loss between policy and expert behaviour over the distribution of tasks, i.e.

Given the above, we define Goal misgeneralization (GMG) as the observed phenomenon in which a system successfully trained to pursue a particular goal in setting fails to generalize to a new setting and instead capably pursues a different goal . A goal in this definition can either be a specific state (static) or a behaviour (dynamic). Note that we use the words "task" and "goal" interchangeably, and will do so for the remainder of this work. A system will be in capable pursuit of a given goal if a metric describing the extent of goal achievement (e.g. success rate) is significantly higher than the corresponding metric for most other goals in . Mathematically, we say GMG happens if

and

We place our definition in between those of Langosco et al. (2022) and Shah et al. (2022), relaxing the former's reliance on RL and Orseau, McGill, and Legg (2018)'s agents and devices framework for simplicity, while focusing on SDM rather than the more general case proposed by the latter, to avoid overly wide characterizations of the phenomenon.

## Afterword

That's all for this excerpt. I hope you found it interesting. I do not claim correctness or strong confidence in the ideas here, but figured it could attract some interest and gather some peer review. This work was carried out between Summer 2022 and October 2023 so may be a bit out of date. As mentioned you can find the rest of the thesis here. Thank you very much for reading!

1. We use the term Sequential Decision Making (SDM) to refer to the field studying problems and approaches wherein an artificial agent interacts with an environment in the process of pursuing and eventually achieving a specific goal (Frankish and Ramsey, 2014). In this context, we envision the agent as acting according to some policy which maps states to actions . States are instantaneous representations of the environment, descriptions of the environment at a given moment. Actions are motions and outputs produced by the agent that may affect the state of the environment. We model the interaction between the agent and the environment as unfolding over discrete time steps. At each time step, the agent observes the state, consults its policy to select an action, and then executes that action. In the next time step, the environment responds by transitioning to a new state, and the loop continues. In other words, the formalism for problems typically studied in Reinforcement Learning and/or Imitation Learning under the Markov Decision Process (MDP) framework (Puterman 2014) ↩︎

2. This generalizes self-proposed tasks, in which the actor is also the requester . ↩︎

New Comment