Goal recognition via model-based and model-free techniques

Abstract

Goal recognition aims at predicting human intentions from a trace of observations. This ability allows people or organizations to anticipate future actions and intervene in a positive (collaborative) or negative (adversarial) way. Goal recognition has been successfully used in many domains, but it has been seldom been used by financial institutions. We claim the techniques are ripe for its wide use in finance-related tasks. The main two approaches to perform goal recognition are model-based (planningbased) and model-free (learning-based). In this paper, we adapt state-ofthe-art learning techniques to goal recognition, and compare model-based and model-free approaches in different domains. We analyze the experimental data to understand the trade-offs of using both types of methods. The experiments show that planning-based approaches are ready for some goal-recognition finance tasks.

Investigating the Combination of Planning-Based and Data-Driven Methods for Goal Recognition

“…We evaluated four different data-driven methods: The NBM presented earlier in Section 5 and three additional data-driven approaches, which were selected following a recent study by Borrajo et al (Borrajo, Gopalakrishnan, & Potluru, 2020). Specifically, we used K-Nearest-Neighbors (KNN) (Russell, 2016, pp.…”

“…First, we used a binary encoding of the planning states, consisting of the state of all planning fluents. Second, following the work of Borrajo et al (Borrajo et al, 2020), we used a vector encoding of the observed action sequence.…”

“…For the LSTM, we also evaluated two different data encodings: A one hot representation of both the observed action sequence and the observed state sequence. Here, we used different vector encodings than for the other data-driven methods because LSTM models, in contrast to the other considered methods, are explicitly designed to handle temporal data sequences which are better captured by one hot data encodings (Borrajo et al, 2020). For both setups, we used the ADAM optimizer with a learning rate of 0.01, a batch size of 32, and 100 epochs.…”

Real-Time Online Goal Recognition in Continuous Domains via Deep Reinforcement Learning

“…Artificial intelligence and machine learning have advanced rapidly, leading to the emergence of a new paradigm for goal recognition based on learning theory. This paradigm can be categorized into two broad groups: model-based and model-free approaches, based on the learning method employed [27,28].…”

“…In contrast, model-free goal recognition based on learning requires no model and only utilizes the observed action sequence and initial state as input. Borrajo et al [27,33] studied goal recognition using XGBoost and LSTM neural networks, which only employ observation sequences without any domain knowledge. However, they trained specific machine learning models for each goal recognition instance and used specific instance datasets for training and testing.…”