Abstract:We study the problem of globally optimising a target variable of an unknown causal graph on which a sequence of soft or hard interventions can be performed. The problem of optimising the target variable associated with a causal graph is formalised as Causal Bayesian Optimisation (CBO). We study the CBO problem under the cumulative regret objective with unknown causal graphs for two settings, namely structural causal models with hard interventions and function networks with soft interventions. We propose Graph Agnostic Causal Bayesian Optimisation (GACBO), an algorithm that actively discovers the causal structure that contributes to achieving optimal rewards. GACBO seeks to balance exploiting the actions that give the best rewards against exploring the causal structures and functions. To the best of our knowledge, our work is the first to study causal Bayesian optimization with cumulative regret objectives in scenarios where the graph is unknown or partially known. We show our proposed algorithm outperforms baselines in simulated experiments and real-world applications.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: how to globally optimize the target variable in the case of an unknown causal graph. Specifically, the researchers explored the Causal Bayesian Optimisation (CBO) problem in two settings: one is to perform hard interventions in Structural Causal Models (SCMs), and the other is to perform soft interventions in Function Networks (FNs).
### Problem Background
In many practical applications, such as in biology, manufacturing, and healthcare, optimizing a target variable usually involves an understanding of causal relationships. However, in the real world, the actual data - generating process is often unknown or only partially known. Therefore, how to effectively optimize in the case of an unknown causal graph becomes an important challenge.
### Specific Problem Description
1. **Unknown Causal Graph**: In many practical applications, the causal graph is usually unknown or not fully known. This makes it difficult to directly apply traditional Causal Bayesian Optimisation methods.
2. **Soft and Hard Interventions**: The researchers considered two types of interventions:
- **Hard Interventions**: Directly set the values of certain nodes, making them independent of their parent nodes.
- **Soft Interventions**: Indirectly affect the values of nodes by controlling variables.
### Solution
To address the above challenges, the researchers proposed the Graph Agnostic Causal Bayesian Optimisation (GACBO) algorithm. The main features of this algorithm are as follows:
- **Actively Discover Causal Structure**: GACBO not only optimizes the target variable but also actively discovers the causal structure that helps to achieve the optimal reward.
- **Balance Exploration and Exploitation**: GACBO explores the causal structure and functions while choosing the best action to obtain the maximum reward.
- **Adapt to Different Intervention Types**: GACBO is applicable to both soft and hard intervention cases.
### Method Overview
GACBO achieves optimization through the following steps:
1. **Initialization**: Start from all possible directed acyclic graphs (DAGs), assuming they have the same prior probability.
2. **Construct Confidence Intervals**: Update the set of possible causal graphs based on the collected data and construct confidence intervals for each node.
3. **Select Intervention**: Use the Upper Confidence Bound (UCB) strategy to select intervention measures, ensuring a balance between exploration and exploitation.
4. **Update Model**: Update the posterior distribution of the Gaussian processes (GPs) according to the new observed data and re - define the confidence intervals.
5. **Repeat Iteration**: Repeat the above steps until the predetermined number of iterations is reached.
### Experimental Results
The researchers evaluated the performance of the GACBO algorithm in multiple synthetic environments (such as Dropwave, Alpine3, Rosenbrock, ToyGraph) and one real - world environment (Epidemiology Graph). The experimental results show that GACBO outperforms other baseline methods in most cases, especially when dealing with unknown causal graphs.
In conclusion, this paper aims to solve the problem of how to effectively optimize the target variable in the case of an unknown causal graph and proposes a novel GACBO algorithm to address this challenge.