Home Tech News AI Paper Summary Microsoft Researchers Introduce a Theoretical Framework Using Variational Bayesian Theory Incorporating a...

Microsoft Researchers Introduce a Theoretical Framework Using Variational Bayesian Theory Incorporating a Bayesian Intention Variable

https://www.microsoft.com/en-us/research/blog/synergizing-habits-and-goals-with-variational-bayes-a-new-framework-for-biological-and-artificial-embodied-agents/

In decision-making, habitual behavior has always been seen as separate from goal-directed behavior. Habitual behaviors are automatic responses, deeply ingrained through experience. Like riding a bike or reaching for your coffee cup in the morning, they required little to no conscious thought. In contrast, goal-directed behavior requires deliberate planning and action to achieve a specific outcome, like finding a new route for the office because of traffic. Due to this separation in both behaviors, the models didn’t capture how habits can influence goals and vice versa.

Microsoft researchers introduce the Bayesian behavior framework to address the traditional division between habitual and goal-directed behaviors in biological and artificial agents. These behaviors are seen as separate entities managed by distinct neural systems: habitual behaviors are fast, automatic, and model-free, while goal-directed behaviors are slow, deliberate, and model-based. The research aims to synergize these two types of behaviors by using variational Bayesian methods.

Current approaches in psychology and neuroscience treat habitual and goal-directed behaviors independently, each relying on different neural mechanisms. Habitual behaviors are quick and automatic but inflexible, whereas goal-directed behaviors are flexible but computationally intensive. To bridge this gap, the researchers introduce a novel Bayesian behavior framework. This framework utilizes variational Bayesian methods to unify these behaviors through a concept called the Bayesian intention variable. This variable represents a dynamic intention that can adjust based on sensory cues (habitual) and specific goals (goal-directed), thereby allowing a seamless transition and interaction between the two behavior types.

The core of the proposed framework involves minimizing the divergence between habitual and goal-directed intentions. This is achieved by combining the habitual and goal-directed intentions using inverse variance-weighted averaging. This unified intention allows agents to leverage the efficiency of habitual behaviors while maintaining the flexibility of goal-directed planning. The framework was tested in vision-based sensorimotor tasks within a T-maze environment, yielding three significant observations:

1. Transition from Goal-Directed to Habitual Behavior: Agents naturally transitioned from slow, goal-directed actions to faster, habitual behaviors through repetitive trials, reducing computational demands on goal-directed processes.

2. Behavior Change After Reward Devaluation: Agents showed resilience in their habitual behaviors despite changes in reward values, reflecting real-world behavioral patterns observed in psychology.

3. Zero-Shot Goal-Directed Planning: Agents efficiently tackled new goals without additional training, demonstrating the framework’s ability to generalize behaviors by leveraging pre-developed habitual skills.

In conclusion, the proposed method presents a significant advancement in understanding and modeling behavior by synergizing habitual and goal-directed actions through a Bayesian framework. This innovative approach not only bridges the gap between these two behavior types but also enhances the efficiency and adaptability of decision-making processes in both biological and artificial agents.


Check out the Paper and blog. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter

Join our Telegram Channel and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our 45k+ ML SubReddit

Pragati Jhunjhunwala is a consulting intern at MarktechPost. She is currently pursuing her B.Tech from the Indian Institute of Technology(IIT), Kharagpur. She is a tech enthusiast and has a keen interest in the scope of software and data science applications. She is always reading about the developments in different field of AI and ML.

🐝 FREE AI WEBINAR: A Synthetic Data Deep Dive (July 30 2024)

Thank You 🙌

X
Exit mobile version