A Bayesian Rule for Adaptive Control based on Causal Interventions

Explaining adaptive behavior is a central problem in artificial intelligence research. Here we formalize adaptive agents as mixture distributions over sequences of inputs and outputs (I/O). Each distribution of the mixture constitutes a `possible world’, but the agent does not know which of the possible worlds it is actually facing. The problem is to adapt the I/O stream in a way that is compatible with the true world. A natural measure of adaptation can be obtained by the Kullback-Leibler (KL) divergence between the I/O distribution of the true world and the I/O distribution expected by the agent that is uncertain about possible worlds. In the case of pure input streams, the Bayesian mixture provides a well-known solution for this problem. We show, however, that in the case of I/O streams this solution breaks down, because outputs are issued by the agent itself and require a different probabilistic syntax as provided by intervention calculus. Based on this calculus, we obtain a Bayesian control rule that allows modeling adaptive behavior with mixture distributions over I/O streams. This rule might allow for a novel approach to adaptive control based on a minimum KL-principle.

💡 Research Summary

The paper addresses the fundamental problem of modeling adaptive behavior in artificial intelligence when an agent interacts with an environment through both inputs and its own outputs. The authors formalize an adaptive agent as a Bayesian mixture over a set of “possible worlds”, each world being a probabilistic model that generates an input‑output (I/O) sequence. The true world is unknown to the agent, and the quality of adaptation is measured by the Kullback‑Leibler (KL) divergence between the I/O distribution of the true world and the I/O distribution expected by the agent, which is uncertain about which world it faces.

For pure input streams (i.e., when the agent only observes data), the standard Bayesian mixture provides a well‑known optimal solution: the posterior mixture minimizes the expected KL divergence, and the agent’s predictions converge to the true distribution as more data arrive. However, the situation changes dramatically when the agent’s outputs are part of the I/O stream. Outputs are not passive observations; they are actions that intervene on the environment. Treating them as ordinary observations leads to a mismatch because the standard mixture does not account for the causal effect of the agent’s own interventions. Consequently, the naïve Bayesian mixture fails to minimize the KL divergence in the I/O setting.

To resolve this, the authors invoke Pearl’s causal intervention calculus. For each possible world they define an interventional conditional distribution (P(I_t \mid do(O_t), \text{world}_i)), where (do(O_t)) denotes the agent’s deliberate choice of output at time (t). This formulation distinguishes between the probability of observing an input given that the agent has forced a particular output, and the ordinary conditional probability that would be used if the output were merely observed. By integrating the do‑operator into the Bayesian update, the mixture correctly reflects the agent’s influence on the environment.

From this causal perspective the authors derive the Bayesian Control Rule (BCR). The rule can be summarized as follows:

1. Initialize prior weights (w_i) over the set of possible worlds.
2. At each time step (t), sample an action (output) (O_t) according to a mixture of the world‑specific policies, weighted by the current (w_i).
3. Observe the resulting input (I_t).
4. Update each weight using the interventional likelihood:
   \