In cryptocurrency networks, transaction-fee revenue serves as the primary financial incentive for miners to participate in the consensus mechanism, securing the network by validating transactions and extending the blockchain. Maximizing this revenue is therefore a key optimization problem for miners. While this has been studied for UTXO blockchains like Bitcoin, where transaction fees are fixed, we focus on Ethereum, where the challenge is significantly greater. On Ethereum, the fee paid by a transaction depends on its execution cost (gas), which can change based on the ordering of preceding transactions in a block. This creates a combinatorial explosion, as miners must select not only a subset of transactions but also their optimal permutation to maximize revenue.

In this work, we present a randomized framework to address this problem. Our approach first uses randomized testing, executing sample permutations of pending transactions to profile their gas usage. From this data, we employ decision trees to learn transaction interdependencies, identifying a small “neighborhood” of transactions that influence each other’s execution costs. These dependencies are then encoded as a set of logical rules that predict gas usage based on local transaction ordering. Finally, we translate these rules and other constraints (e.g., block gas limit, nonce ordering) into an integer linear programming (ILP) instance, which we solve to find a block composition that maximizes total tip revenue. Our experimental results demonstrate significant gains: our method outperforms real-world Ethereum miners by an average of 73.45 percent per block, which corresponds to roughly 63 million USD per annum.