At Cenacle we study the effects of group dynamics when asymmetric robots are involved, where each robot’s capabilities are different, in terms of their hardware capabilities and computational power.
Identifying recurring patterns is a key ability for machine learning to work correctly, and forms the base for autonomous robots to build models of their environments accurately. When the learning is happening in a single robot, the collected knowledge is all stored at a single place (usually inside the robot’s internal persistent memory structure) accessible to it at all times.
When multiple robots come together to explore a complex environment, taking part in the learning process as a cluster, the collected knowledge is spread across the cluster of machines and the model-building process becomes complicated. The robots might have collected information that, individually, is sparse, but collectively forms a complete model. When there is no single centralized designated repository for all the collected information to be processed centrally, as happens in the autonomous machinery scenarios, where there are constraints on power, communication and computation resources, the robots should learn to work with sparse models of truth that is incomplete most of the time, but accurate to the extent available. Such sparse distribution of knowledge across multiple autonomous robots, forces the robots to cooperate with each other to achieve their goals with the limited knowledge each possesses.
When centralized computation is possible, techniques such as streaming variational approximations [Broderick et al., 2013], parallel stochastic gradient descent [Niu et al., 2011], the Map-Reduce framework [Dean and Ghemawat, 2004], database-inspired concurrency control [Pan et al., 2013], and message passing on graphical models [Gonzalez et al., 2009] etc. can be used for distributed learning. This takes advantage of consistent global model shared by all robots, with which they can make local updates without the concern of generating conflicts unknown to each other [Campbell et al., 2014]. When such central computation is not possible, however, issues such as asynchronous communications, lack of consistent global state, unreliable communications etc. pose major challenge and lead to inconsistencies in the model processed and possessed by each robot.
Distributed Ledgers enable autonomous intelligent machinery to perform collaborative problem solving
At Cenacle we study these challenges and propose solutions required to arrive at collaborative planning and goal oriented work division across multiple autonomous robots. One significant aspect of the study is to consider the effects of group dynamics when asymmetric robots are involved, where each robot’s capabilities are different, in terms of their hardware capabilities and computational power. The primary aim is to focus on:
- How autonomous intelligent machinery can make collaborative problem solving possible, and
- What are the consequent social structures that evolve due to such collaboration.
Collaborative problem solving, at the outset, requires the participating entities to have a common goal to start with. However, when the entities are autonomous, formulating a common goal in itself is a major challenge, due to the absence of a central command structure. This become further complex, when each of the entities have their own individual goals, and further some of those individual goals are conflicting with each other.
The participating intelligent machinery, aka robots, should learn to arrive at a common goal, and further:
- Learn to perform goal subdivision: when the common goal is complex, or cannot be performed by a single robot, it should be divided into individual tasks that can be executed by more than one robot, either in parallel or in sequence.
- Task planning and distribution: The distribution of tasks to robots for execution requires correct sequence planning based on the inter-task dependencies, and individual robot’s capabilities.
- Knowledge gaining and sharing: When a task to be executed by a robot requires certain knowledge about the task, or the environment, which the robot does not possess by itself, it should either work to attain the knowledge either by trying to acquire it by itself, or from some other robot which has it [GK Palem, 2005].
Such collaboration among the participating entities soon results in the formation of social structures centered around:
- Asymmetric capabilities: When each robot’s capabilities differ, in terms of hardware capabilities or processing power (such as few robots are equipped with fast processing GPUs that can perform image detection and analysis tasks at high speeds, while some robots are very light weight flying drones that are equipped with only cameras without processing system), each robot tries to benefit from the others on the capabilities they lack on their own through the cooperative efforts
- Synergy models: When the robots are autonomous, without any central command structure, and given their asymmetry of capabilities, the need for cooperation will have to established through various synergy models, such as, Trade off one’s capabilities
- for additional knowledge, or
- Gain priority over limited resources (such as power recharge units etc.).
- Get additional priority for own goals over others’ etc.
- Sociotechnical systems: When the participating robots are owned by multiple individual humans, each coming with its own restrictions and individual goals, performing common tasks in such a way that does not contradict the restrictions and helps the individual goals, results in a complex socio-technical dynamic between humans and the participating robots.
Autonomous intelligent machinery capable of patrolling and collaborative 3D architectural reconstruction is a new species emerging based on Blockchain distributed ledgers and Artificial intelligence.
- T. Campbell, Jonathan How. Approximate Decentralized Bayesian Inference. 30th Conference on Uncertainty in Artificial Intelligence, UAI 2014, Quebec City, Quebec, Canada, July 23-27, 2014. p.1-10
- T. Broderick, N. Boyd, A. Wibisono, A. C. Wilson, and M. I. Jordan. Streaming variational bayes. In Advances in Neural Information Procesing Systems 26, 2013.
- F. Niu, B. Recht, C. R´e, and S. J. Wright. Hogwild: A lockfree approach to parallelizing stochastic gradient descent. In Advances in Neural Information Processing Systems 24, 2011.
- J. Dean and S. Ghemawat. Mapreduce: Simplified data processing on large clusters. In 6th Symposium on Operating Systems Design and Implementation, 2004.
- X. Pan, J. Gonzalez, S. Jegelka, T. Broderick, and M. I. Jordan. Optimistic concurrency control for distributed unsupervised learning. In Advances in Neural Information Processing Systems 26, 2013.
- J. E. Gonzalez, Y. Low, and C. Guestrin. Residual splash for optimally parallelizing belief propagation. In Proceedings of the 12th International Conference on Artificial Intelligence and Statistics, 2009.
- GK Palem, Data-dependencies and Learning in Artificial Systems, In Proc. of Approaches and Applications of Inductive Programming, pages 69–78, 2005.