Robust Player Plan Recognition in Digital Games with Multi-Task Multi-Label Learning

Abstract

Plan recognition is a key component of player modeling. Player plan recognition focuses on modeling how and when players select goals and formulate action sequences to achieve their goals during gameplay. By occasionally asking players to describe their plans, it is possible to devise robust plan recognition models that jointly reason about player goals and action sequences in coordination with player input. In this work, we present a player plan recognition framework that leverages data from player interactions with a planning support tool embedded in an educational game for middle school science education, CRYSTAL ISLAND. Players are prompted to use the planning tool to describe their goals and planned actions in CRYSTAL ISLAND. We use this data to devise data-driven player plan recognition models using multi-label multi-task learning. Specifically, we compare single-task and multi-task learning approaches for both goal prediction and action sequence prediction. Results indicate that multi-task learning yields significant benefits for action sequence prediction. Additionally, we find that incorporating automated detectors of plan completion in plan recognition models improves predictive performance in both tasks.

Language Model-Based Player Goal Recognition in Open World Digital Games

“…Although a player may have more than one goal or plan in mind when establishing them, an action at a moment is usually taken for a single goal or plan. In prior work (Goslen et al 2022a), they removed completed goals and plans from among multilabels to partially alleviate this problem, but the issue remains while multi-labels are maintained until only a single label remains. As a result, many actions performed for a single goal or plan can become unclear under multi-labels.…”

“…To validate our goal recognition approach, we utilized the dataset and labels provided by prior work using CRYSTAL ISLAND as a testbed (Goslen et al 2022a). The dataset which is an open-world digital game designed for middle school science education (Figure 5).…”

“…Assumptions: Table 1 summarizes the assumptions of our goal recognition problem in CRYSTAL ISLAND, following the assumption criteria of Masters and Vered (2021) CRYSTAL ISLAND, the environment is continuous since the player's behavior is open and infinite, but the continuous actions and states were discretized to high-level concepts and locations in the provided data, and thus it is discrete. When interacting with players, the game environment is online, but we utilized offline data collected from a prior study (Goslen et al 2022a). This particular version of game is also static and deterministic because the environment's rules do not change over time, and there is no randomness in state transition.…”

“…Labeling: Generally, labeling of player goals and plans is non-trivial since it requires knowing the players' intention during gameplay. Prior work (Goslen et al 2022a) encouraged players to externalize their goals and plans through the use of a planning support tool. The planning support tool of-fered players a choice of 20 goals, categorized into 5 goal classes, and 55 available abstract actions that players could use to create plans for their goals.…”

“…To evaluate our work, we use data from an open-world educational game, CRYSTAL ISLAND, used in prior research on goal recognition (Goslen et al 2022a) and employ T5 (Raffel et al 2020) as our base language model. The results show that the T5-based models outperform non-linguistic machine learning approaches, and that T5 with TCA further improves the goal recognition performance.…”