Emergent Mind

Read to Play (R2-Play): Decision Transformer with Multimodal Game Instruction

(2402.04154)
Published Feb 6, 2024 in cs.AI and cs.LG

Abstract

Developing a generalist agent is a longstanding objective in artificial intelligence. Previous efforts utilizing extensive offline datasets from various tasks demonstrate remarkable performance in multitasking scenarios within Reinforcement Learning. However, these works encounter challenges in extending their capabilities to new tasks. Recent approaches integrate textual guidance or visual trajectory into decision networks to provide task-specific contextual cues, representing a promising direction. However, it is observed that relying solely on textual guidance or visual trajectory is insufficient for accurately conveying the contextual information of tasks. This paper explores enhanced forms of task guidance for agents, enabling them to comprehend gameplay instructions, thereby facilitating a "read-to-play" capability. Drawing inspiration from the success of multimodal instruction tuning in visual tasks, we treat the visual-based RL task as a long-horizon vision task and construct a set of multimodal game instructions to incorporate instruction tuning into a decision transformer. Experimental results demonstrate that incorporating multimodal game instructions significantly enhances the decision transformer's multitasking and generalization capabilities.

Overview

  • Introduces multimodal game instructions (MGI) integration into Decision Transformers (DTs) for improved multitasking and generalization in Reinforcement Learning (RL).

  • Describes the novel Decision Transformer with Game Instruction (DTGI) model, which combines textual and visual instructions and a unique design, SHyperGenerator, for enhanced adaptability.

  • Presents empirical evidence showing that DTs with MGI outperform those with singular modal instructions, particularly in unseen gaming environments.

  • Envisions a future where the integration of multimodal instructions in AI and LLMs can lead to superior performance across diverse tasks and challenges.

Introduction

In the realm of AI, developing generalist agents that exhibit adeptness across diverse tasks has been a long-standing objective. Reinforcement Learning (RL) approaches, empowered by extensive offline datasets, have demonstrated remarkable multitasking capabilities. Nevertheless, these models often grapple with the adaptation to unfamiliar tasks due to the limitations in accessing task-specific knowledge and contextual information. While recent advancements have attempted to surmount these barriers with textual or visual guidance, the effectiveness of singular modal guidance remains inadequate for providing comprehensive contextual task understanding. This paper posits that the utilization of multimodal game instructions could significantly elevate the performance of Decision Transformers (DTs) by offering enriched contextual cues, thereby facilitating superior multitasking and generalization.

Multimodal Game Instruction: A New Frontier

The inception of multimodal game instructions (MGI) marks a pivotal advancement in the pursuit of crafting more versatile and adaptable RL agents. Drawing inspiration from the efficacy of multimodal instruction tuning in visual task performance enhancement, this study pioneers the integration of these instructions into the Decision Transformer framework, thus birthing the Decision Transformer with Game Instruction (DTGI). This novel configuration not only leverages the collective strengths of textual and visual instructions but also introduces a unique design, SHyperGenerator, to enable between-task knowledge sharing, further augmenting the model's adaptability to unseen gaming environments.

Compelling Experimental Insights

The empirical evaluations underscore the significant improvements ushered in by the incorporation of MGIs into DTs. The findings distinctly illustrate that:

  • DTs equipped with MGI markedly outperform those facilitated by singular modal instructions, underscoring the superior comprehensive nature of multimodal contextual information.

  • The adaptability and performance of DTs in unseen games escalate noticeably with the integration of MGIs, highlighting the model's enhanced generalization capabilities.

  • A larger dataset size and a greater diversity of training games proportionately improve both in-distribution (ID) and out-of-distribution (OOD) performance, suggesting the multifaceted benefits of MGI across various scales of data availability.

The Future Trajectory of AI and LLMs

This research provides a compelling demonstration of how multimodal instructions can revolutionize decision-making processes in RL-driven models. The integration of MGI into DT not only reflects a significant leap in the development of generalist agents but also paves the way for future explorations into the realms of AI and LLMs. The potential for a generalized multimodal instruction framework looms on the horizon, promising enhancements in performance across not just vision-based tasks but also in the broader landscape of AI challenges.

In summation, the integration of MGIs into DTs heralds a new era in the evolution of AI, where the symbiosis of multimodal cues and decision-making processes unfurls new dimensions of learning, adaptability, and task execution. The journey forward is ripe with possibilities for extending this innovative approach across various domains, further solidifying the foundation for more intelligent and versatile AI agents.

Get summaries of trending AI/ML papers delivered straight to your inbox

Unsubscribe anytime.

YouTube
References
  1. The arcade learning environment: An evaluation platform for general agents. J. Artif. Intell. Res., 47:253–279.
  2. Language models are few-shot learners. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H., editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
  3. Open-world multi-task control through goal-aware representation learning and adaptive horizon prediction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 13734–13744. IEEE.
  4. GROOT: Learning to Follow Instructions by Watching Gameplay Videos
  5. AutoAgents: A Framework for Automatic Agent Generation
  6. Decision transformer: Reinforcement learning via sequence modeling. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W., editors, Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 15084–15097.
  7. Delta Tuning: A Comprehensive Study of Parameter Efficient Methods for Pre-trained Language Models
  8. Palm-e: An embodied multimodal language model. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J., editors, International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202 of Proceedings of Machine Learning Research, pages 8469–8488. PMLR.
  9. Minedojo: Building open-ended embodied agents with internet-scale knowledge. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A., editors, Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9
  10. ChessGPT: Bridging Policy Learning and Language Modeling
  11. HiFi: High-information attention heads hold for parameter-efficient model adaptation. In Rogers, A., Boyd-Graber, J., and Okazaki, N., editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8521–8537, Toronto, Canada. Association for Computational Linguistics.
  12. Hypernetworks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net.
  13. Towards a unified view of parameter-efficient transfer learning. In International Conference on Learning Representations.
  14. Parameter-efficient transfer learning for NLP. In Chaudhuri, K. and Salakhutdinov, R., editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 2790–2799. PMLR.
  15. Inner monologue: Embodied reasoning through planning with language models. In Liu, K., Kulic, D., and Ichnowski, J., editors, Conference on Robot Learning, CoRL 2022, 14-18 December 2022, Auckland, New Zealand, volume 205 of Proceedings of Machine Learning Research, pages 1769–1782. PMLR.
  16. Hyperdecoders: Instance-specific decoders for multi-task NLP. In Goldberg, Y., Kozareva, Z., and Zhang, Y., editors, Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1715–1730, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.
  17. Deep Reinforcement Learning with Multitask Episodic Memory Based on Task-Conditioned Hypernetwork
  18. Compacter: Efficient low-rank hypercomplex adapter layers. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W., editors, Advances in Neural Information Processing Systems, volume 34, pages 1022–1035. Curran Associates, Inc.
  19. Parameter-efficient multi-task fine-tuning for transformers via shared hypernetworks. In Zong, C., Xia, F., Li, W., and Navigli, R., editors, Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 565–576, Online. Association for Computational Linguistics.
  20. Multi-game decision transformers. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A., editors, Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9
  21. The power of scale for parameter-efficient prompt tuning. In Moens, M.-F., Huang, X., Specia, L., and Yih, S. W.-t., editors, Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045–3059, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
  22. Otter: A Multi-Modal Model with In-Context Instruction Tuning
  23. Parameter-efficient fine-tuning without introducing new latency. In Rogers, A., Boyd-Graber, J., and Okazaki, N., editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 4242–4260, Toronto, Canada. Association for Computational Linguistics.
  24. Visual Instruction Tuning
  25. Zero-shot reward specification via grounded natural language. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvári, C., Niu, G., and Sabato, S., editors, International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 14743–14752. PMLR.
  26. Human-level control through deep reinforcement learning. Nat., 518(7540):529–533.
  27. OpenAI (2021). Chatgpt: A large-scale generative model for open-domain chat. https://github.com/openai/gpt-3.

  28. Training language models to follow instructions with human feedback. In NeurIPS.
  29. Conceptual reinforcement learning for language-conditioned tasks. In Williams, B., Chen, Y., and Neville, J., editors, Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 9426–9434. AAAI Press.
  30. AdapterFusion: Non-destructive task composition for transfer learning. In Merlo, P., Tiedemann, J., and Tsarfaty, R., editors, Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, pages 487–503, Online. Association for Computational Linguistics.
  31. Learning transferable visual models from natural language supervision. In Meila, M. and Zhang, T., editors, Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, volume 139 of Proceedings of Machine Learning Research, pages 8748–8763. PMLR.
  32. Language models are unsupervised multitask learners. Github.
  33. Generalization to New Sequential Decision Making Tasks with In-Context Learning
  34. A generalist agent. Trans. Mach. Learn. Res.
  35. Feature importance estimation with self-attention networks. In Giacomo, G. D., Catalá, A., Dilkina, B., Milano, M., Barro, S., Bugarín, A., and Lang, J., editors, ECAI 2020 - 24th European Conference on Artificial Intelligence, 29 August-8 September 2020, Santiago de Compostela, Spain, August 29 - September 8, 2020 - Including 10th Conference on Prestigious Applications of Artificial Intelligence (PAIS 2020), volume 325 of Frontiers in Artificial Intelligence and Applications, pages 1491–1498. IOS Press.
  36. Generative Multimodal Models are In-Context Learners
  37. Aligning Large Multimodal Models with Factually Augmented RLHF
  38. Describe, Explain, Plan and Select: Interactive Planning with Large Language Models Enables Open-World Multi-Task Agents
  39. Interactive Natural Language Processing
  40. Feature importance ranking for deep learning. In Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., and Lin, H., editors, Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.
  41. Read and Reap the Rewards: Learning to Play Atari with the Help of Instruction Manuals
  42. Hyper-decision transformer for efficient online policy adaptation. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net.
  43. Multiinstruct: Improving multi-modal zero-shot learning via instruction tuning. In Rogers, A., Boyd-Graber, J. L., and Okazaki, N., editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 11445–11465. Association for Computational Linguistics.
  44. One network, many masks: Towards more parameter-efficient transfer learning. In Rogers, A., Boyd-Graber, J., and Okazaki, N., editors, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7564–7580, Toronto, Canada. Association for Computational Linguistics.
  45. LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention
  46. Instruction Tuning for Large Language Models: A Survey
  47. Prototype-based HyperAdapter for sample-efficient multi-task tuning. In Bouamor, H., Pino, J., and Bali, K., editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 4603–4615, Singapore. Association for Computational Linguistics.

Show All 47