Infinite Ends from Finite Samples: Open-Ended Goal Inference as Top-Down Bayesian Filtering of Bottom-Up Proposals (2407.16770v1)

Abstract: The space of human goals is tremendously vast; and yet, from just a few moments of watching a scene or reading a story, we seem to spontaneously infer a range of plausible motivations for the people and characters involved. What explains this remarkable capacity for intuiting other agents' goals, despite the infinitude of ends they might pursue? And how does this cohere with our understanding of other people as approximately rational agents? In this paper, we introduce a sequential Monte Carlo model of open-ended goal inference, which combines top-down Bayesian inverse planning with bottom-up sampling based on the statistics of co-occurring subgoals. By proposing goal hypotheses related to the subgoals achieved by an agent, our model rapidly generates plausible goals without exhaustive search, then filters out goals that would be irrational given the actions taken so far. We validate this model in a goal inference task called Block Words, where participants try to guess the word that someone is stacking out of lettered blocks. In comparison to both heuristic bottom-up guessing and exact Bayesian inference over hundreds of goals, our model better predicts the mean, variance, efficiency, and resource rationality of human goal inferences, achieving similar accuracy to the exact model at a fraction of the cognitive cost, while also explaining garden-path effects that arise from misleading bottom-up cues. Our experiments thus highlight the importance of uniting top-down and bottom-up models for explaining the speed, accuracy, and generality of human theory-of-mind.

Open-Ended Goal Inference via Top-Down Bayesian Filtering and Bottom-Up Proposals

In the paper "Infinite Ends from Finite Samples: Open-Ended Goal Inference as Top-Down Bayesian Filtering of Bottom-Up Proposals," the authors Tan Zhi-Xuan et al. propose a computational model to address the problem of inferring human goals in an open-ended setting. Existing models largely operate under the assumption of a fixed and finite set of possible goals, yet human goal inference capabilities seem to significantly transcend these limitations. This paper advances the theoretical and algorithmic understanding of human goal inference in open-ended and complex scenarios.

The proposed model utilizes a Sequential Monte Carlo (SMC) approach to combine top-down Bayesian inverse planning with bottom-up heuristic sampling. This approach enables the model to quickly generate plausible goal hypotheses and subsequently filter them through a rationality framework. This bifurcated approach harmonizes the efficiency and flexibility of human goal inference capabilities.

Methodology

The authors introduce their model in a rigorous computational framework. The core elements include:

• Goal Prior: $g \sim P(g)$
• Online Planning: $\pi_t \sim P(\pi_t | s_{t-1}, \pi_{t-1}, g)$
• Action Selection: $a_t \sim P(a_t | s_{t-1}, \pi_t)$
• State Transition: $s_t \sim P(s_t | s_{t-1}, a_t)$

These elements together form a generative model to perform Bayesian inverse planning. Given the computational challenges posed by large goal spaces, typical of open-ended settings, the authors propose a novel integration. They leverage the familiarity of humans with the statistics of their environment to hypothesize goals rapidly via bottom-up proposals.

The bottom-up sampling relies on the hypothesis that humans, based on environmental cues and knowledge of common subgoals, generate and test goal hypotheses. Upon hypothesizing a goal, the model utilizes goal filtering based on top-down rationality constraints, forming a two-stage hypothesis: generate plausibly and then evaluate rigorously.

Experimental Setup

To validate their model, the authors employed the "Block Words" task, where participants observed sequences of actions made by an agent stacking lettered blocks to form a word. Participants then inferred potential goal words at several judgment points within the sequence. The experimental conditions were systematically varied to test different types of inference challenges, such as bottom-up friendly scenarios, irrational alternatives, garden paths, and uncommon words.

Results

The open-ended SMC model demonstrated superior performance in approximating human inferences, achieving higher similarity to human goal distributions, particularly in the intricate scenarios like irrational alternatives. These results substantiate the value of combining bottom-up sampling for hypothesis generation with top-down Bayesian filtering for hypothesis evaluation.

Specifically, the empirical results indicated:

• Accuracy: The open-ended SMC model showed comparable accuracy to exact Bayesian inference in simpler tasks but also outperformed pure bottom-up heuristics in complex scenarios.
• Human Similarity: The model displayed higher intersection over union (IoU) with human participants' inferences across all conditions, particularly excelling in scenarios requiring rational evaluation of goals.
• Algorithmic Properties: The model replicated the variance and sample efficiency observed in human participants, suggesting a plausible cognitive mechanism.
• Resource Rationality: The open-ended model balanced computational efficiency with high accuracy, supported by sensitivity analysis showing robust performance under different parameter configurations.

Implications and Future Work

The paper has significant theoretical and practical implications. Theoretically, it bridges the gap between fixed-goal and open-ended goal inference models, demonstrating that human-like goal inference can be both computationally tractable and behaviorally accurate. Practically, this model has potential applications in areas requiring human-AI interaction and collaborative systems, enhancing the ability of AI to anticipate and align with human goals dynamically.

Future research could further refine the model's bottom-up proposal mechanism using more advanced LLMs, and explore adaptive algorithms that better mimic human cognitive processes like goal rejuvenation and selective forgetting. Integrating these enhancements could yield more robust and scalable open-ended goal inference systems.

Conclusion

The presented model represents a significant advancement in understanding the computational mechanisms underpinning human goal inference in open-ended scenarios. By effectively combining the strengths of bottom-up sampling and top-down Bayesian filtering, the model not only aligns closely with human reasoning patterns but also offers practical utility in complex real-world tasks. The findings underscore the potential for further developments in AI systems that can better understand and predict human goals in dynamic and open-ended environments.