Heuristic search algorithms rely on a heuristic function, $h$, to guide search for planning. The aim of such a heuristic function is to produce a quick-to-compute estimate of the true cost-to-goal, $h^*$, for any given state $s$.
A well-known property of heuristic search based algorithms like A* or IDA* is that if the heuristic never overestimates the true cost-to-goal - that is, $h(s) \leq h^*(s)$ - then the plans produced by these algorithms is guaranteed to be optimal. Such a heuristic is called an admissible heuristic.
Unfortunately, crafting strong admissible heuristics is difficult, often requiring expert domain-specific knowledge and high memory resources.
An alternative approach is to learn heuristics from data using machine learning algorithms. For example, consider the popular 15-puzzle. The aim of the 15-puzzle is to reach a goal state from some start state by sliding a blank tile in a direction and swapping that blank tile with the adjacent number in that direction.
Now, suppose we have a set of optimal plans for a set of 15-puzzle tasks, where each plan is for a 15-puzzle task with a different start state. Then it is possible to use these plans as training data for a supervised learning algorithm, such as a neural network, to learn a heuristic. Since supervised learning algorithms generalise to unseen data, such heuristics can then be applied to new, previously unseen tasks i.e. 15-puzzle tasks with different start states...
Congratulations to Kale-ab Tessera whose work, Learning compact, general purpose neural network architectures, received the "Best Poster" award at the 2019 Deep Learning Indaba.
He wins a trip to Vancouver, Canada in December, where he will attend NeurIPS 2019.
We are proud to sponsor the Women in Computational Science Research event at the University of the Witwatersrand on 8 August 2019.
In this work, we investigate ways an agent can combine existing skills to create novel ones in a manner that is both principled and optimal. We find that by constraining the reward function and transition dynamics, skills composition can be achieved in both entropy-regularised and standard RL. Our approach allows an agent to generate new skills without further learning, and can be applied to high-dimensional environments and deep RL methods.
In this post, we look to answer the following question: given a set of existing skills, can we compose them to produce novel behaviours? For example, imagine we have learned skills like running and jumping. Can we build more complex skills
by simply combining them in interesting ways? Of course, it is likely that there are many ad-hoc ways to do this, but it would be really nice if we could do it in a way that is both provably correct and that requires no further learning.
To illustrate what we're after, we use pretrained skills generated by DeepMimic [Peng, 2018]. As we can see, in the first two animations below our agent has learned to backflip and spin. But if we try to combine these skills together (in this case, by taking the maximum of the policy network output), we get complete bupkis! The question is: is there some way we can make this work?
Attempting to compose two skills - backflips and spins - results in an agent that can injure itself optimally.
Unsurprisingly, the answer is "yes". More...
We are excited that Sicelukwanda Zwane has been interviewed on the This Week in Machine Learning (TWiML) podcast on his current research on safer exploration in deep reinforcement learning using action priors.
Through Social Cobots: Robots with Human-Like Collaboration Skills
This study is a part of a collaboration with DAI-Labor of TU Berlin, Germany. We envision the future of collaborative robots (cobots) in industry through their fully autonomous human-like collaboration with human partners.
Our research aims to address the question: "How do we build cobots with human-like natural collaboration skills?". Existing intention-aware planning approaches often make the assumptions that a human collaborator's actions are always relevant to the collaborated task and that the human always accepts the cobot's assistance when offered.
We believe that these assumptions are a significant barrier against having social cobots in the real world. In reality, a human's dynamic desires and emotional states could result in stochastic human intentions, especially in repeated tasks. A cobot with these assumptions may misinterpret the human actions, which may result in intrusive and unreasonable robot behaviors (e.g. a human gazing at an object might be interpreted as she needs it, yet behind this gaze, she could be evaluating to take it herself or thinking of something irrelevant to the task like staring into space).
Our goal is to offer a new model design as an alternative to the conventional intention-aware models by removing these assumptions. The result is our novel robot decision-making model, a partially observable Markov decision process (POMDP), that is capable of handling these dynamic...